They spent more than $30B over 10 years before turning an operating profit in 2023.

But once it did, the company became a free cash flow machine:

Just as the company was finally hitting its stride, it now faces an existential threat with Autonomous Vehicles. The assumed risk is simple — a company like Waymo can create its own app, sidestepping Uber, and provide the service at a lower cost than Uber can with its human drivers.

But in our view, this is simplistic thinking that doesn't account for why Uber lost all that money during the building phase — ride-hailing is not a global network effect. If you are a marketplace like Amazon, you can reliably service a few states or even the entire country with a massive warehouse once you have built it up. But for Uber, the network you have built in New York isn't useful to someone trying to get a cab in Boston. In the U.S. alone, Uber has spent nearly a decade building 8,000+ micro-markets to match supply with demand.

From an Autonomous Vehicle (AV) developer's perspective, it’s much better (operationally and financially) to plug into the existing Uber ecosystem than to build a system independently (more on this later).

How Uber makes money

To get a sense of how big Uber has become, they have 200mm+ active users and enabled ~13B trips (rides + delivery) in 2025 alone. It also runs a B2B delivery arm, but it accounts for a small fraction of the overall business and adds minimal value.

Let’s go from the least to most important segment.

Freight:

Uber Freight is a digital freight brokerage that matches shippers with truck carriers. Rides and freight book differently. In rides, one clearing price is visible to both sides, and Uber takes a disclosed cut. In freight, shippers and carriers negotiate prices separately. Uber contracts as principal, committing a price to each side, so it books the entire gross booking as revenue.

Delivery:

Launched in 2014, it quickly scaled to become the second-largest local delivery business globally ex-China, with $85B in total GBV in 2025, closely trailing the largest player, DoorDash, at $100B. Last month, Uber announced a cash offer for Delivery Hero at an equity value of $14.8 billion. Once the acquisition closes with the necessary approvals, Uber will be the largest delivery business, with ~$135 billion in GBV.

Rides :

This is Uber’s largest and most profitable business, with $100B in gross bookings value (GBV) and $7.5B in EBITDA in 2025. That profitability shows up in its high take rate of ~30% (revenue retained after driver payouts) and EBITDA of 7.6% of GBV.

Rides GBV has sustained 20% growth for 15 years because Uber keeps innovating on its product; it now operates in 10+ categories, from motos to luxury cars, tailored to local demand. Product innovation, rising demand density, and falling cost per mile for EVs (and eventually AVs) are enabling Uber to expand into suburban markets and increase overall market penetration.

Every Uber ride breaks down into three major costs.

1. Driver income. The median Uber driver grosses $21.18 an hour, or about $12.50 per trip. That is roughly 70% of the rider fare. After fuel and maintenance, the driver keeps just over 50%.

2. Insurance. This is the second highest cost. Most US states mandate high-value commercial coverage for rideshare trips, which can run up to 30 times a normal policy. Depending on the state, that is 10% to 30% of the fare. We assume about $2 per trip.

3. Vehicle operating cost. AAA puts fuel and maintenance at about 25 cents per mile. That works out to roughly 15% of the fare.

Autonomous Vehicles | Landscape, Unit Economics & Adoption

AV Landscape:

AVs face a high bar to launch: a robotaxi must be meaningfully safer than a human driver. But the bar is fixed, unlike GPUs, where performance keeps doubling every few years, helping the likes of NVIDIA stay ahead of the competition. AVs have a finish line on safety and reliability.

More than a dozen companies globally are racing to reach this finish line. Waymo got there first in the US, but several well-funded companies (Zoox, Tesla, etc.) are in advanced trials with launch timelines in the next few years.

Additionally, Mobileye and NVIDIA are building the tech stack to make AVs plug-and-play technology that any automobile company can license or buy. Stellantis and Volkswagen are already taking this route to build AVs.

With so many players racing to build a technology that has a finish line, we believe the AV tech stack will be commoditized.

As the technology gets commoditized, AV hardware is also getting cheaper. Waymo’s Jaguar I-PACE fleet cost $150,000–200,000 per unit, ~$75,000 for the vehicle and ~$125,000 for autonomous hardware. Its replacement, the Zeekr-built Ojai, costs ~$60k, with a ~$38k base vehicle and AV hardware under $20k, using 42% fewer sensors.

AV scaling is supported by declining cost per mile as hardware costs fall and AV companies compete. AV companies’ success now depends on their ability to scale and capture existing demand — and Uber is in the best position to provide that demand.

AV unit economics & utilization:

AVs carry a large share of operating costs as fixed overhead — fleet management, maintenance, depreciation, etc. To cover these costs, AVs need high utilization rates. As trips scale from 10 per day to 30 per day, fixed costs can drop from ~30% of revenue to ~10% of revenue, improving AV profit margins.

Currently, AVs break even at ~20+ trips a day. AVs launched on Uber’s network are already operating in this zone of mid-to-high 20s to low 30s trips per vehicle per day.

Uber has reported that AVs on its platform generate 30% more trips than on proprietary apps. Consumer ETAs are also 25% lower on Uber, incentivizing riders to keep coming back (the classic platform effect). The only way to break Uber’s grip on demand is to repeat what Uber itself did: heavy consumer discounts and enough supply to hold ETAs down.

This utilization gap on Uber versus proprietary apps drives Uber’s take rate. 30% higher utilization means 30% more revenue on the same AV fleet. Even after Uber takes a 15% cut, AVs still come out ahead.

Uber's take rate is bounded by the incremental utilization it generates for the AV. When an AV player is fully dependent on Uber, Uber can likely charge a flat ~20–30% take rate.

China offers a preview of the central role aggregators play in driving demand and utilization for AVs. T3 Chuxing, China’s #3 ride-hailing operator, saw its orders from aggregator platforms climb from 61.5% to 77.5% to 85.9% over three years. T3 owns and operates its cars, but demand increasingly flows through aggregators.

AV adoption rate & Market size:

Waymo’s cost per mile in the US is estimated at ~$2 in 2025, comparable to that of human drivers on Uber. By 2030, AV cost per mile is expected to fall below $1 as hardware costs decline and fleet management gains economies of scale.

AVs are under 1% of the total cab market today, but within their operating markets they have captured up to a 20% share. Uber has commented on increasing market share in these markets, showing early signs of how AV platforms drive adoption and benefit from it.

As cost per mile declines and AV fleets expand with multiple competing players, AVs are expected to reach ~20% of the US cab market by 2030. With cabs at ~2% market share of personal vehicle miles, lower-cost AVs have room to grow the cab market itself, not just take share within it.

How AV’s will affect Uber’s Effective take rate

Uber’s rides business today has a headline take rate of ~30%, but its incremental contribution margin is only ~10%. If we look deeper at where the other 20% of take rate goes, most of it falls into insurance and driver incentive programs during peak demand. These two components consume 15% of the headline take rate, leaving Uber’s effective take rate at ~15%.

While AVs might carry a slightly lower headline take rate, Uber doesn’t have to pay peak-demand incentives, and insurance costs will be minimal because AVs prove safer & AV operators are likely to absorb this cost themselves as part of fleet negotiations. Thus, even if Uber accepts a lower take rate of 15% on AV rides, it remains competitive with the effective take rate of human-driven trips.

Uber Model and Market Sentiment Rating

To summarize our findings:

• Robotaxis won’t be a winner-take-all market and will most probably end up a split race with multiple players clearing the bar.

• From the supplier side, they will look to maximize utilization, and from the consumer side, users will look for the fastest taxis — both of which Uber is perfectly positioned to service.

• Directionally, AVs should improve Uber's net take rate. And the deflationary pricing cuts both ways: lower prices expand the market by pulling more users into AV rides.

Read more

No paywall on this one. Share it with a friend :)

The elephant in the room in every AI discussion is whether hyperscaler capex is sustainable. The bulls are happy as the numbers keep getting revised higher every quarter.

On the other end, bears look at the $1T number and call it unsustainable.

What surprised us is that we couldn’t find anyone who actually worked through the numbers to figure out who’s right. While certainty is impossible in investing, just like building a DCF model, working through the numbers gives us a plausible estimate of what conditions are needed to sustain the current spending.

This is crucial as nearly 80% of the Market Sentiment portfolio and ~50% of the S&P 500 is directly exposed to AI or AI-adjacent spending. Any pullback here will bring a brutal drawdown.

So get a cup of coffee and let’s dig into what is probably the most important investing question of the decade! To answer this, we need to model

1. What goes into a datacenter, and how long it lasts.

2. What the new capacity earns, and for how long.

3. What’s the running cost, and how would the margins evolve over time?

Join 46,000 other investors to make sure you don’t miss our next briefing.

Subscribe now

What goes into an AI datacenter?

Today, a 1GW datacenter is expected to cost ~$40B. This capex splits into two categories defined by asset life.

1. Chips — compute and memory

Traditionally, chips last 5 to 6 years, after which a more efficient generation makes them uneconomic to run, and reliability begins to drop. While cloud companies are trying to extend this life by moving older chips to lower-intensity inference workloads, for our model we are assuming a 5-year life (more on this later).

CoreWeave’s A100s (a 2020 release) are still fully booked, and a batch of Nvidia H100s (2022) that came off contract was immediately re-let at ~95% of the original price. CoreWeave CEO Mike further noted (Q1 2026) sustained pricing across the older fleet as inference pushes demand up for the older-generation devices.

2. Datacenter shell — building, power equipment, networking

The useful life of buildings runs up to 30 years. Google commented on its recent earnings call that ~60% of capex went to servers and ~40% to datacenters and networking equipment. The shell’s share of the mix has come down from ~50%+ a few years ago as semis have become more expensive.

Given rising prices with each new GPU generation, we assume a semis-to-shell capex ratio of 2:1. So for every $100B spent on a new datacenter, $67B goes to chips and $33B goes to the shell.

We take the top-of-the-range shell life because the shell is chip-agnostic. The building, power, and cooling serve successive generations of chips without major changes. Also, the permitted sites are one of the scarcest assets in the buildout, and the owners would want to push as much mileage out of them as possible.

This capex drives the entire AI semis complex. Microsoft, Amazon, and Google plan to spend ~$600B on capex in 2026, versus ~$300B in 2025. This spending gets redirected one way or the other into the chip side (Nvidia, AMD, TSMC, SK Hynix, etc.) and the shell side (Vertiv, GE, Schneider Electric, Caterpillar, etc.).

What the capacity earns?

A datacenter earns by renting compute. Two variables set that revenue: how much of the capacity gets rented, and the rate per GPU-hour.

The first bit is easy for now: all data points to the fact that capacity is sold before it is even built. AWS backlog reached $496B in Q2 2026, and Amazon has said the lion’s share of its 2027 capacity is already reserved. Every other hyperscaler shows the same trend. Microsoft closed June 2026 with $678B of commercial RPO, up 84% year over year; Google Cloud’s backlog hit $514B in Q2 2026, up more than $50B in a single quarter.

The rate per GPU-hour is the tricky bit, as it keeps falling as the GPU ages. For example, the one-year committed pricing for H100s has fallen from ~$3.00 to ~$1.80-2.40 per GPU-hour over three years.

The average H100 rental rate over its first three years was ~$2.25/hr. At 90% utilization, that is ~$18k of revenue per GPU per year. A $100B datacenter puts ~$67B into chips at the 2:1 ratio, which at ~$32k per H100 buys ~2.1 million GPUs.

So roughly, a $100B datacenter will generate ~$37B of revenue a year over its first three years.

What it costs to run?

A datacenter has two major operating costs: energy and maintenance.

Electricity: Datacenters are measured in power; as the name suggests, a 1 GW datacenter consumes 1 GW of power. But there are efficiency losses between power drawn from the grid and power actually used by the compute, measured as PUE (Power Usage Effectiveness). Datacenters have an estimated PUE of 1.14, so a facility with 1 GW of IT load draws 1.14 GW from the grid. Assuming a compute utilization rate of ~90%, total annual consumption comes to 1 GW × 90% × 1.14 × 8,760 hours = 8,987 GWh (~9.0 TWh).

US average industrial power cost is 8.66 cents per kWh as of April 2026, or $86,600 per GWh. Datacenters are built where power is cheapest, with low-cost states running below that; for our model, we assume the US industrial average. This gives a total power cost of 8,987 GWh × $86,600 per GWh = $778mm, or approximately $800mm per year for a 1 GW datacenter at a 90% utilization rate.

Maintenance: Maintenance costs are primarily spread across property taxes, labor, and repair work. As per Epoch AI, average maintenance for a 1 GW datacenter stands at ~$300mm across these three categories.

So, all in, the annual cost of running a 1 GW datacenter stands at ~$1.1B with $800mm in energy at today’s prices and $300mm in maintenance. As per a Morgan Stanley report, this can go as high as $2B depending on power costs and property taxes at the location. For our model, we will be conservative and take $2B as the annual cost of operating a 1 GW datacenter.

A 1 GW datacenter is estimated to cost ~$40B to build. With $100B in capex, we can build 2.5 GW, leading to total operating costs of ~$5B per year.

The Model

Bringing it all together, we can now model the cash generation of a $100B datacenter.

• At a 2:1 semis-to-shell ratio, $67B goes to chips and $33B to the shell.

• At 90% rental occupancy, the datacenter will generate $37B in annual revenue for the first 3 years and then $24B for years four and five (assuming a 1/3rd decrease in GPU rental pricing).

• Operating costs hold at ~5B per year across 5 years.

Now assume $100B is invested in datacenters every year, and let’s see what steady-state revenue it will generate.

Every $100B of capex buys chips with a five-year life. For the first three years, that cohort generates $37B of revenue annually. In years four and five, revenue drops to $25B as pricing falls with chip age — the decay we saw earlier.

Now spend $100B every year and run it forward. By year five you’re operating five vintages at once: three young cohorts earning $37B each and two aging cohorts earning $25B each.

(3 × $37B) + (2 × $25B) = $161B

That’s steady state. From year five onward, one cohort retires as a new one comes online, so revenue holds at ~$160B while capex stays at $100B.

So a $100B annual capex will generate ~$160B of annual revenue over the long run — a revenue-to-capex multiple of ~1.6x.

Based on our model at this 1.6x multiple and $5B annual costs, a cloud company generates an IRR of ~21%.

It is worth remembering our conservatism here. We modeled a 5-year semiconductor life, while CoreWeave currently runs up to 6+ years; hyperscalers also have better negotiating power and run a mix of GPUs and more efficient ASICs, both of which support lower operating costs. Adjusting for these, the IRR should land even higher.

Share

Benchmarking against real data

The model implies chip spend is recovered in roughly three years. Now let’s check that against what an H100 purchased in 2023 would actually have earned, based on its purchase price and realized rental rates. An H100 cost an estimated $25–40k in 2023.

Based on SemiAnalysis’s rental price tracker, the average H100 rental rate over its first three years (2023-26) was ~$2.25/hr, peaking at ~$3.00/hr in the early period. Assuming 80% utilization and a 75% cash margin, that generates ~$11.8k of annual cash flow, implying a payback of ~2.7 years. With cloud companies already locking in capacity for 2027-28, realized utilization may run above our 80% base case.

_{H100 ROI Model | Source: Market Sentiment Research}

Morgan Stanley’s unit economics model for a 1GW datacenter built on Nvidia’s latest GB300 chips assumes a materially shorter payback in the current pricing environment. The model implies $23B of semis capex, generating annual cash profit of ~$15B at the low end, leading to a payback of ~1.5 years on the chip spend.

For reference, if we take Morgan Stanley’s pricing (payback period of 1.5 years) for the first 2 years and a 1/3rd price decline every year after that and plug it into our model, we will have an IRR of ~30%. This is how attractive current compute pricing is.

Can revenue get there?

The top three cloud companies are expected to spend ~$600B on capex in 2026, against ~$300B in 2025. Applying the 1.6x revenue-to-capex multiple to that base, they need to generate ~$960B in cloud revenue to sustain the current level of spend.

Today’s combined run rate is ~$380B, growing at 40%+ and still accelerating.

So rather than debating whether the current semiconductor drawdown is a correction or the start of a bubble bursting, the question can be reduced to something simpler: can cloud companies grow their revenue to $960B?

We think they have a reasonable chance. Every major cloud is accelerating on an already large base:

• AWS is at a ~$169B run rate, with growth accelerating to 37% from 17.5% a year ago.

• Google Cloud has crossed a ~$99B run rate, with growth jumping to 82% from ~32%.

• Azure is past $100B and still accelerating, to 43% from ~39%.

At the current ~40% growth rate, cloud revenue reaches a ~$1T run rate in three years and ~$1.5T in four. Amazon CEO Andrew Jassy has said he expects AWS alone will likely scale to a $1T revenue business over time.

Further, when thinking about their capex, it helps to remember internal capex needs for R&D and first-party products like Search, YouTube, Office and Prime, which would require spending regardless of AI for cloud services. Cloud companies spent ~$100B of capex in 2022, before AI began scaling for consumers.

All in, we think the 2026 capex of ~$600B is sustainable as revenue scales. We will revisit this next year to test whether the Street’s ~$1T estimate for 2027 holds up against actual revenue and bookings growth.

Where it can break

While current growth rates and bookings are great, we see three major risks to cloud demand.

1. The compute demand does not show up

Everything above assumes enterprises and consumers keep using more AI. Usage is growing today, and enterprise spend is rising with it.

Now that the token-maxing era has passed, companies like Coinbase and Uber are optimizing AI budgets by placing monthly spend limits per employee.

But while some firms trim employee AI credits, aggregate spend keeps climbing as AI penetrates internal workflows. In BCG’s survey of 2,600 C-suite executives, only 6% plan to cut AI spend, while average budgets are expected to double in 2026 versus 2025

2. One big player blows up, and there goes the backlog

Over 40% of cloud backlog traces to two model labs: an estimated ~$700B to OpenAI and ~$330B to Anthropic.

Anthropic last reported a ~$47B run rate as of May 2026, up from ~$9B at the end of 2025, with roughly 80% of revenue from enterprises. Enterprise revenue is stickier and more sustainably monetized, which makes the $330B commitment look better covered.

OpenAI’s run rate is at $25B since February 2026 compared to 2025 revenue of $13.1B and a $20.9B operating loss. It filed a confidential S-1 in June 2026 but has since leaned toward delaying the listing to 2027. We view the OpenAI backlog as carrying more risk.

The right question, though, is what actually happens if OpenAI disappears tomorrow. As a ChatGPT user, I won’t stop using AI. I will switch products, and that platform’s compute demand will absorb OpenAI’s.

3. Decline in demand from VC-funded AI companies

The third risk is a pullback in demand from VC-backed AI startups. Venture funding rose to ~$650B in 1H26 from ~$350B in 2023. Look deeper, though, and a large share concentrates in the top five companies, essentially the AI labs. Exclude those and 1H26 funding is ~$220B, against ~$300B for all of 2023, a difficult year for venture capital.

We therefore don’t see venture investment as a primary driver of AI compute demand. A decline in venture funding would matter at the margin, but ex-labs venture investment has not scaled meaningfully with AI in the first place.

None of this makes it a sure thing. Every bank in the street underwrote the fiber buildout in the dot-com bubble. It worked till it didn’t.

But with this, we have a sneak peek into the math that the hyperscalers and the banks that are underwriting the multi-billion-dollar bonds are looking at. While we are working with public data, they would have access to at least 100x more data points flowing directly from the datacenters.

This is also exactly why we are comfortable positioning MS portfolio across these three categories and have conviction to hold through the market drawdowns.

ASICs (~20% of portfolio)

ASICs enable hyperscalers to improve profitability on the current revenue base and unlock new use cases by pushing down token cost.

Memory (~33% of portfolio)

Memory has seen a sharp correction because of unwinding leveraged trades. With sustained capex of $600B and a large part of it going to inference rather than training, we see memory as a capacity-constrained market for the next one to two years and remain bullish in the near term.

Advanced Packaging (~25% of portfolio)

This category contains manufacturing lines (TSMC, Intel, Amkor) whose revenue is directly linked to chip demand; hence, we remain bullish on them. Capital equipment firms (LRCX, AMAT, KLAC) depend on growth in capex, not sustainability of capex. Sustained capex means no need for new manufacturing lines, which will directly hit their revenue growth. We are least bullish on these firms and will trim our positions as capex guidance peaks at $1T in 2027.

If you made it till here, I would love to hear what you think:

Leave a comment

That’s it for now. As you can see, hundreds of hours of research go into each of these reports. Please consider upgrading your subscription to support our work and to get access to all the reports.

Subscribe now

Judging by DMs we received last week, a lot of you are running portfolios wildly outside your risk tolerance. While we cannot recommend custom portfolio strategies, a simple trick I have learned over more than half a decade is to build a sleep-well portfolio.

The setup is simple: your portfolio should be structured in a way that you can sleep through the market open without missing a beat. Most of us are not fund managers, and there is no good reason to check your portfolio every hour (and it's detrimental to your happiness irrespective of how the portfolio performs). If you are all-in, or worse, 4x levered like Leopold on names that have gone up 2-5x in the last year, you are asking for trouble. Having a reasonable exposure will also help you think rationally, which is exactly what we were doing a few days back (with the thesis playing out literally the next day).

With that, let’s get into the big developments over the past month, how our portfolio performed, and how we are positioning over the next few months.

Overall, it was quite a bad month for AI bulls with the SOXX semiconductor ETF down 21% and memory names taking a beating (DRAM was down ~30% last month).

We see this correction as unwinding of leverage and portfolio concentration that’s been building in semis. Korea’s leveraged ETFs had ~$10B AuM at the beginning of 2026. By June it ballooned to ~$50B with SK Hynix-linked leveraged products alone at ~$23 billion.

On July 16, South Korea’s Financial Services Commission announced stricter regulatory rules for single-stock leveraged ETFs, raising the minimum margin requirement from 10 million Korean Won to 30 million Korean Won and limiting purchases to a maximum of 20 shares per person per transaction.

During the stock price decline phase, the mandatory end-of-day rebalancing mechanism of leveraged ETFs triggered programmatic automatic selling, instantly creating a sharp selloff. That day, SK Hynix fell another 11%, Samsung slumped over 8%, and the wave of panic quickly swept through Europe and the US. Leopold’s $45 billion AI fund blowing up last week added fuel to the fire.

While the correction can be largely attributed to normal market volatility amplified by leverage, there were quite a few material events:

The launch of Kimi K3

Moonshot released Kimi K3 on July 16, the largest open-source model with 2.8 trillion parameters. AI capability benchmarks have placed it on par with OpenAI 5.5 and Opus 4.8.

Kimi follows a dozen open-source, low-cost models China has released since DeepSeek R1. In our view, AI adoption gains far outweigh open-source AI efficiency gains, as reflected in Nvidia’s sharp rebound post the first DeepSeek event.

But companies are currently in the token-maxxing era. Every employee, whether they need it or not, has access to frontier models with a nearly unlimited AI budget. There will come a time when companies see the AI bill and ask if it’s really providing the ROI. This does not mean they will stop using AI. They will just optimize it better.

There’s no need to use Claude Fable 5 to rewrite your email or update your deck for the 20th time. Open source models can do the same at 1/100th of the cost. Serving intelligence becomes a race to the lowest cost per token, the same way Japanese carmakers won the US market by delivering most of the performance at a much lower price.

CXMT had a monster IPO

CXMT is a Chinese memory company holding 8% of global DRAM capacity. CXMT listed on Shanghai on July 27, raising $8.6B in Asia’s biggest IPO this year. It popped 466% on debut, valuing it at ~$500B+ as one of the most valuable China-listed companies.

China’s semiconductor capability buildout is well known; however, it lags the leading edge by a few nodes, and CXMT itself doesn’t have HBM capability.

New memory capacity by Korean players

Samsung and SK announced investment of ₩800T (~$516B) in four new fabs to double memory capacity within five years. This is the most significant news for memory names with questions on disciplined supply. However, new capacity comes online only in 2H27–28, keeping memory in tight supply for the next 4Qs.

These plants are backed by major contracts for Korean memory names. SK has long-term contracts of ~$700B ($500B with Nvidia, Microsoft, Anthropic) while Samsung has ~$200B (Broadcom MOU).

While the Capex addition is real, the quantum of Capex should be taken with a pinch of salt, as these announcements were made as part of a South Korean government super-industrial plan and not a corporate strategy event. Historical precedent shows that memory companies will delay Capex if they don’t see supportive pricing (2022-23 delayed wafer starts). Thus, we see this capacity addition plan having no impact for the next 1-2 years, and the supply-demand dynamic at that point will be tested based on AI adoption.

MS Portfolio Update

Read more

FYI — Live tracker is only available to paid subscribers and can be accessed at the end of this report.

First things first.

Judging by DMs we received last week, a lot of you are running portfolios wildly outside your risk tolerance. While we cannot recommend custom portfolio strategies, a simple trick I have learned over more than half a decade is to build a sleep-well portfolio.

The setup is simple: your portfolio should be structured in a way that you can sleep through the market open without missing a beat. Most of us are not fund managers, and there is no good reason to check your portfolio every hour (and it’s detrimental to your happiness irrespective of how the portfolio performs). If you are all-in, or worse, 4x levered like Leopold on names that have gone up 2-5x in the last year, you are asking for trouble. Having a reasonable exposure will also help you think rationally, which is exactly what we were doing a few days back (with the thesis playing out literally the next day).

With that, let’s get into the big developments over the past month, how our portfolio performed, and how we are positioning over the next few months.

Overall, it was quite a bad month for AI bulls with the SOXX semiconductor ETF down 21% and memory names taking a beating (DRAM was down ~30% last month).

We see this correction as unwinding of leverage and portfolio concentration that’s been building in semis. Korea’s leveraged ETFs had ~$10B AuM at the beginning of 2026. By June it ballooned to ~$50B with SK Hynix-linked leveraged products alone at ~$23 billion.

On July 16, South Korea’s Financial Services Commission announced stricter regulatory rules for single-stock leveraged ETFs, raising the minimum margin requirement from 10 million Korean Won to 30 million Korean Won and limiting purchases to a maximum of 20 shares per person per transaction.

During the stock price decline phase, the mandatory end-of-day rebalancing mechanism of leveraged ETFs triggered programmatic automatic selling, instantly creating a sharp selloff. That day, SK Hynix fell another 11%, Samsung slumped over 8%, and the wave of panic quickly swept through Europe and the US. Leopold’s $45 billion AI fund blowing up last week added fuel to the fire.

While the correction can be largely attributed to normal market volatility amplified by leverage, there were quite a few material events:

The launch of Kimi K3

Moonshot released Kimi K3 on July 16, the largest open-source model with 2.8 trillion parameters. AI capability benchmarks have placed it on par with OpenAI 5.5 and Opus 4.8.

Kimi follows a dozen open-source, low-cost models China has released since DeepSeek R1. In our view, AI adoption gains far outweigh open-source AI efficiency gains, as reflected in Nvidia’s sharp rebound post the first DeepSeek event.

But companies are currently in the token-maxxing era. Every employee, whether they need it or not, has access to frontier models with a nearly unlimited AI budget. There will come a time when companies see the AI bill and ask if it’s really providing the ROI. This does not mean they will stop using AI. They will just optimize it better.

There’s no need to use Claude Fable 5 to rewrite your email or update your deck for the 20th time. Open source models can do the same at 1/100th of the cost. Serving intelligence becomes a race to the lowest cost per token, the same way Japanese carmakers won the US market by delivering most of the performance at a much lower price.

CXMT had a monster IPO

CXMT is a Chinese memory company holding 8% of global DRAM capacity. CXMT listed on Shanghai on July 27, raising $8.6B in Asia’s biggest IPO this year. It popped 466% on debut, valuing it at ~$500B+ as one of the most valuable China-listed companies.

China’s semiconductor capability buildout is well known; however, it lags the leading edge by a few nodes, and CXMT itself doesn’t have HBM capability.

New memory capacity by Korean players

Samsung and SK announced investment of ₩800T (~$516B) in four new fabs to double memory capacity within five years. This is the most significant news for memory names with questions on disciplined supply. However, new capacity comes online only in 2H27–28, keeping memory in tight supply for the next 4Qs.

These plants are backed by major contracts for Korean memory names. SK has long-term contracts of ~$700B ($500B with Nvidia, Microsoft, Anthropic) while Samsung has ~$200B (Broadcom MOU).

While the Capex addition is real, the quantum of Capex should be taken with a pinch of salt, as these announcements were made as part of a South Korean government super-industrial plan and not a corporate strategy event. Historical precedent shows that memory companies will delay Capex if they don’t see supportive pricing (2022-23 delayed wafer starts). Thus, we see this capacity addition plan having no impact for the next 1-2 years, and the supply-demand dynamic at that point will be tested based on AI adoption.

MS Portfolio Update

Read more

The jump in performance per watt was incredible. M chips did the same tasks Intel chips did at roughly a quarter of the power draw. The result: the M1 Air managed 15 to 18 hours of real use versus barely 8 hours on the Intel Air with the same battery capacity.

Apple achieved this with custom silicon. Intel could never build this chip for Apple because Intel had to serve everyone. Its processors carried decades of baggage and had to work in every Windows machine from a $300 Dell to a gaming rig, so they could not be optimized for one laptop or one operating system.

Apple’s decade long investment into custom silicon paid off — Mac revenue which was stagnated at ~$25B from 2017 to 2019 jumped 60% and touched ~$40B by 2022 before normalizing to $33B for the latest year.

This is not about laptops; It is that owning your silicon converts directly into performance per watt, and performance per watt converts directly into a product nobody else can build.

Nowhere is this now more important than AI. Our estimates show that for every $100 of inference cost, hyperscalers currently hand roughly half to Nvidia. This eye-watering 85% gross margin Nvidia charges for its AI chips is currently the single largest tax on the AI buildout, and it is a tax you can escape by designing your own silicon.

Market Sentiment dives deep into the bottlenecks of the AI world. Join 46,000 other investors to make sure you don’t miss our next briefing.

Subscribe now

What is an ASIC?

ASIC stands for Application Specific Integrated Circuit. It is a chip designed by one company for its own specific use case, rather than sold on the open market. Examples include:

• The Apple M series chips we just discussed

• Google’s TPU AI processor

• Amazon’s Trainium 3 AI processor

• Nokia’s ReefShark base station processor

If you are wondering how these differ from Nvidia or AMD GPUs: a GPU is a Swiss Army knife whereas an ASIC is a scalpel.

Nvidia builds its GPUs to run anything. Training, inference, any model architecture for text, vision, protein or physics, in any precision format. That flexibility comes at a price. A meaningful share of every GPU die is spent on circuitry your specific workload will never touch.

An ASIC flips the trade. You strip out everything except what your workload needs and pour the entire silicon budget into exactly that. Bitcoin mining chips are the purest example. They compute one hashing function billions of times a second and can do literally nothing else. Not even a calculator app would run on one.

The AI versions apply the same logic to neural networks. The result is 2 to 3x better performance per watt and real savings on silicon cost: Google's TPU7x matches NVIDIA's B200 on raw compute at roughly a third of the price: ~$13,000 per TPU vs $35,000.

Artificial Analysis benchmarked open source Llama models across hardware. The results below are clear: Google's TPUs squeeze roughly 2x more tokens out of each TFLOP than any Nvidia chip.

Your Margin Is My Opportunity

With AI going mainstream, the nature of AI compute is changing. Inference is starting to take a larger component of overall compute. Inference workloads are predictable and repetitive: exactly what custom chips are built for. Projections show that just the inference demand is the fastest growing segment at a 35% CAGR and will need 90GW of data center by 2030.

The catch with ASICs is that they are expensive to build. A leading-edge custom chip costs $500M+ in design, IP licensing, and tape-outs before a single unit ships. The math only works at scale. You need enough deployment volume that the per-chip savings outrun the upfront design cost.

Every hyperscaler has now crossed that threshold and is building their own chips.

Google has been at it the longest, and its TPUs are the most mature ASICs in production. Amazon is not far behind with Trainium and Inferentia, which we believe will boost AWS returns in the coming years. Meta is the newest entrant: its MTIA chips already run recommendation inference internally, and in February it signed a multibillion-dollar deal to lease Google’s TPUs, with talks underway to install them in Meta’s own data centers from 2027.

The race to serve models at the lowest cost

In our view, companies are currently in the token-maxxing era. Every employee, whether they need it or not, has access to frontier models with a nearly unlimited AI budget.

This will not last, for two reasons.

First, the C-suite is starting to demand ROI for the spend. Uber burned through its entire 2026 AI budget in four months on coding tools, and its COO now admits it is very hard to draw a line between token consumption and useful product output. Microsoft reportedly canceled most of its direct Claude Code licenses this month and moved engineers to GitHub Copilot.

Second, not every task needs a frontier model. Writing an email response or updating a deck does not require Claude Fable 5. An open source model can do the same work at a tenth of the cost or less, and as open models get smarter, the share of workloads running on them will keep expanding.

Note that this does not mean AI spend shrinks. Usage will keep exploding while the price per token collapses, so the pie grows even as every workload gets cost-optimized. What it means is that serving intelligence becomes a race to the lowest cost per token, the same way Japanese carmakers won the US market by delivering most of the performance at a much better price.

If that is the race, hyperscalers have no option but ASICs.

Investing in Accelerating ASIC Adoption

The custom ASIC market is projected to grow from 8% of AI accelerators today to 19% by 2033, a $118 billion segment compounding at 27%. Given how fast the open source models are improving, we believe this projection might be understated.

We have identified 5 companies that sit directly in this flow, and we are backing them with ~20% of the MS portfolio.

Read more

Quite a few of you have joined us over the past week (we were #5 in the Rising Finance Bestsellers) and here’s a guide on what we do at Market Sentiment and how to use the subscription effectively.

Investing in asymmetric opportunities

There are only two ways to make consistent alpha in the market:

1. Be early

2. Be contrarian

It’s futile to look at a company that’s covered by 20 analysts and expect to find something that all of them have missed. Our focus at Market Sentiment is to be early to trends that are going to define the next few years.

Whether you like it or not, that trend right now is AI. We are pumping a mind-bending amount of capital into compute. To put things in perspective (all figures adjusted for inflation):

• Every month, we are spending more on AI infrastructure than the total cost of the Manhattan Project.

• Every quarter, we now allocate more to AI than the entire 14-year Apollo program.

• Just last year, we spent more than what it took to build our entire Interstate highway system over three decades.

AI is the consensus trade now. But you can still be early to what it depends on. Capital at this scale never flows evenly. The railroad mania of the 1800s bankrupted most railroads but made fortunes for the companies selling steel and rights of way. Every infrastructure boom in history worked this way.

The AI buildout will be no different. The hyperscalers spending hundreds of billions will not be where the returns concentrate — that trade is long over. The returns will concentrate at the bottlenecks that these companies can’t sidestep. Our focus is to find them before the market prices them in.

Thematic Coverage

We have built five thematic baskets to date. Three of the five have already re-rated. Energy and Agentic AI have not, which is why they are worth a closer look right now.

1. Energy — AI needs energy. A lot of it. And more importantly, it needs it now. The power demand for data centers is expected to grow 10x the historical generation capacity growth. This is the one basket that’s underperformed (4% vs 15% for the S&P 500), and the reason was a simple mistake: we underestimated how fast the price of power would turn political, and over-allocated to companies exposed to electricity prices. The fix is just as simple: strip those out, and the rest of the basket handily beats the market (more on this below).

2. Optics — We were one of the first to cover the Optics trade last year. While the tech behind is complex, our thesis was simple: As models become more powerful and AI pushes data transfer speeds higher, the industry has no choice but to adopt optical interconnects. The basket has now rerated and is up 75% vs. 16% for QQQ.

3. Advanced Packaging — Once transistor shrinks stopped being cheap, the industry stopped chasing smaller nodes and started gluing chips together. Advanced packaging became the way to keep Moore's law alive, and the bottleneck moved from the wafer to the package. The basket is now up 102% vs 15% for QQQ.

4. Memory — Everyone knows memory stocks are cyclical and have burned investors before. But that’s the wrong question. The right question is whether we’re in a structural shift — and the data says yes. We made memory our highest allocation (34%) and were handily rewarded with a 134% weighted return (we bought the dip during the Iran war).

5. Agentic AI — We are at the inflection point for Agentic AI. Most AI conversations right now are about coding assistants, but that’s really only useful for the ~25 million developers out there. With Claude now generally available in Microsoft Office, AI can meaningfully support the 1.4 billion knowledge workers who spend their days buried in email, spreadsheets, and jumping between a dozen tools.

MS Portfolio

Our portfolio is where we put capital behind these emerging trends. Below you'll find every position, its current buy/sell/hold rating, and all the changes we've made along the way. If you're just starting out, begin by gradually allocating to our buy-rated names — these are the companies we still consider undervalued.

Market Sentiment Portfolio

Two important notes:

1. Many of you have DM’d us asking how you should allocate. As much as we’d like to help, we’re not financial advisors and can’t give customized allocation advice. What we can share is exactly what’s in the sheet: our positions, allocations, and buy/sell/hold ratings, plus how we think about each of them in our deep dives.

2. Starting Monday, this sheet moves behind the paywall. Our full portfolio, including positions, allocations, and ratings, will be available exclusively to paid subscribers. If you’ve been on the fence, this week is a good time to upgrade.

Subscribe now

What’s coming?

Which brings us to the good stuff. There are two big themes that we have been tracking for a while and will be sharing with you over the next few weeks.

1. ASICs

ASICs (Application Specific Integrated Circuits) are projected to overtake GPU unit shipments by next year.

ASICs are chips custom designed for one specific workload. Because these chips are tailored at the transistor level, they are highly efficient at the task its designed to do.

Google’s TPU7x matches NVIDIA’s B200 on raw compute at roughly a third of the price: ~$13,000 per TPU vs $35,000.

The catch here is that ASIC development is a multi-year, multi-billion-dollar project that only hyperscalers like Google and Amazon can afford. But that’s exactly who is building them. As model winners emerge and applications standardize, ASICs lead the next wave of compute growth.

Here’s how we look at the ASIC supply chain: design tools (Cadence, Synopsys), design services (AVGO, Marvell), and licensed IP blocks (Arm, Aspeed).

The real prize sits with the designers. Complex AI ASIC design is effectively an AVGO–Marvell duopoly. This lets them charge fixed development fees plus a variable price per chip.

Given ~95% unit growth and per-chip pricing power, AVGO and Marvell earnings should see superior earnings growth over the next 2–3 years. AVGO expects its AI semiconductor revenue to grow from approximately $20 billion in FY25 to over $100 billion in FY27 primarily driven by these ASICs.

When your data-center investments run into hundreds of billions, spending a few billion on a custom ASIC is a no-brainer.

2. Energy

One mistake we made with the energy basket: we expected power prices to keep rising as demand scaled. Demand did scale. But power prices became a political problem rather than a supply-demand equation. In January, the governors of all 13 states in the PJM grid region, along with the Trump administration, jointly declared that power prices needed to be reined in and that data centers should pay for the grid costs they create. The price ceiling on capacity auctions was then extended to 2030, taking the best-case scenario we had modeled for these companies off the table.

The fix is to separate the demand thesis from the price thesis. The demand thesis was right and remains right. The price thesis is now capped by regulators until 2030.

So we are restructuring the basket around companies that get paid on volume, not price. That means the companies selling the physical building blocks of the grid: gas turbines, transformers, switchgear, high-voltage cables, backup power, and cooling systems, along with the contractors who install all of it. A transformer sold to an Amazon data center faces no price ceiling, because voters don’t pay for it.

That’s it for now. Leave a comment if you have an idea/theme that you think we all should be looking at.

Leave a comment

Disclaimer: Market Sentiment work is provided for informational purposes only, is intended solely for readers in the United States, and should not be construed as legal, business, investment, or tax advice. You should always do your own research.

What’s more interesting is the shift in mainstream sentiment. It has moved from “AI investment is setting cash on fire” to “maybe these trillion-dollar companies aren’t spending for nothing.”

Take the chart below from the FT:

While estimates should always be taken with a bucket load of salt, it’s worth digging into what they are projecting:

1. The CapEx is expected to plateau (not drop) — Once you get past 2027, the CapEx is expected to stop growing at 70 to 80% a year that we are seeing now and settle into a maintenance plus refresh level.

2. Revenue will catch up quickly — If inference, cloud, and ad monetization compound while CapEx flattens, operating cash flow keeps climbing.

3. Finally, CapEx is paid upfront but counts against profits gradually in the form of depreciation over the life cycle of the asset. As the accumulated asset base grows, the depreciation add-back in the FCF bridge gets very large.

If this plays out and free cash flow explodes, the multiples will re-rate immediately. These names now carry a discount due to CapEx uncertainty. When quarterly prints start confirming that capex is flattening and cash is converting, the uncertainty discount will lift and the multiple will expand.

While all 3 are projections, we have some signs of revenue catching up. Bloomberg reported last week that AI revenue has now caught up with the quarterly CapEx depreciation. If the revenue trend holds while capex plateaus, it confirms that the trillions hyperscalers poured into AI capex were worth it.

Global AI sales, excluding China, reached $25 billion in the first quarter of 2026, exceeding the industry’s estimated $21 billion in depreciation costs tied to investments in data centers and chips for the second consecutive quarter. — Bloomberg

Market Sentiment dives deep into the bottlenecks of the AI world. Join 46,000 other investors to make sure you don’t miss our next briefing.

Subscribe now

That brings us to the MS portfolio. We are more or less in line with where we were at the beginning of last month, with the MS portfolio up 75% YTD. Memory continues to be our strongest performer, with SK Hynix, Samsung, and Micron all performing exceptionally well.

MU -6.25%↓ trades around $1,050, roughly 8–9x forward earnings — versus Nvidia’s ~30x trailing multiple, for a company that’s just as levered to AI capex. The gap exists because investors still expect memory prices to crash as they did before.

The Strategic Capacity Agreements (SCAs) are Micron’s direct answer to that skepticism: floor pricing through 2030 for a significant share of its supply, signed by the same hyper-scalers and automakers. We believe the market is still not pricing this long-term structural shift in memory demand and profitability for the next 3-5 years.

As we highlighted last week, we will continue to hold these names and expect them to rise further as the market digests the impact of long-term contracts and becomes less concerned about the cyclicality of memory stocks.

Update on Optics

Our Optics basket was a bet on AI data centers hitting the “Copper Wall”, a physical limit beyond which copper cables cannot transmit data fast enough over a given distance.

Existing high-speed copper cables can transmit data at 112 Gbps over a distance of 2 meters. The next generation of servers is projected to run at twice that speed, and at that point traditional passive copper will no longer reliably carry a signal across a server rack.

The basket was built intentionally to get exposure to both approaches being developed to overcome this problem:

1. Add signal boosters to existing copper data cables (Active Electrical Cables)

2. Convert the electrical signal into an optical signal, achieving near-perfect transition (Co-Packaged Optics – CPO)

Since its launch (Dec’25), our Optics basket is now up 72% vs 14% of QQQ. This outperformance was driven by AEC volumes scaling faster than planned: in the most recent quarter, Credo Technology Group reported ~150% YoY revenue growth and Astera Labs ~90%.

While AEC volume scaled, our optical bets also got several validating events:

• Nvidia announced a multibillion‑dollar purchase commitment for Co-Packaged Optics along with a $2B investment in Coherent.

• Arista Networks also launched XPO-MSA, a switch that uses removable optical connectors. It’s a stepping stone toward denser optical connections while the industry works toward fully integrated optics (CPO).

AEC companies benefit from mature products and are expected to grow by ~60%-80% through 2027, with a strong outlook through 2030. CPO companies are expected to grow ~30%-40% in 2027 as their initial product scales, with potential for better growth over the long run.

Post this strong rally, most portfolio companies in our basket are trading above 100 TTM P/E. Given this, we are updating the optics basket from Buy to Hold.

MS Portfolio Holdings Update

Read more

No paywall on this one. Share it with a friend :)

This one’s a bit different from our usual reports. I just saw an incredible stat from the Nomura research team, and it felt like a good time to revisit our research on Leveraged ETFs from two years ago.

Leveraged ETFs have grown nearly tenfold in the past decade. AUM now sits at an all-time high — this surge is a key driver of the rising volatility we are seeing in markets. The more concerning part is what sits underneath it. Most of this money is retail chasing performance in semiconductors. Over $50B in leveraged ETF AUM is now concentrated in semiconductor funds.

Investors buy 2X and 3X leveraged ETFs expecting 2X and 3X returns. It rarely plays out that way. The risk of capital loss runs far higher than intuition suggests. Here is what happened to the 2X-leveraged MicroStrategy ETF, a fund with over $1B in AUM.

Market Sentiment dives deep into the bottlenecks of the AI world. Join 46,000 other investors to make sure you don’t miss our next briefing.

Subscribe now

To understand what’s going on, we have to first understand how leverage works in investing:

Let’s say that you had $10K in your brokerage account. You have a high-conviction stock you think is going up and want to invest more than $10K in it. You can take another $10K as a loan from your brokerage, staking your entire position as collateral.

Assume the stock went up 10% over the next month, and you exit the position.

This is the magic of leverage. You used a 2:1 leverage on your investment and doubled your return1

What’s important to realize is that this works the other way, too — If the stock went down 10% and you had to exit the position, the 2:1 leverage would literally double your losses.

Having a large amount of leverage is like driving a car with a dagger on the steering wheel pointed at your heart. If you do that, you will be a better driver. There will be fewer accidents but when they happen, they will be fatal - Warren Buffett

Understanding Leveraged ETFs

It’s vital to understand how leveraged ETFs work — class action lawsuits were filed (and later thrown out) because investors did not understand how they are structured.

For example, let’s take a quick look at the TQQQ fact sheet:

The fund is only trying to 3x the daily return of QQQ and not the long-term return.

To understand how the index gains and index volatility affect the total return, consider this — Let’s say you invested $100 into QQQ and TQQQ. On day 1, the index went up 10%. On day 2, the index went down 10%.

Here’s how both funds would have performed2

• Day 1:

  • Index (QQQ): 100 x (1 + 10%) = $110

  • 3x Leveraged Index (TQQQ): 100 x (1 + 3 x 10%) = $130

• Day 2:

  • Index (QQQ): 110 x (1 - 10%) = $99

  • 3x Leveraged Index (TQQQ): 130 x (1 - 3 x 10%) = $91

• Overall Return (after 2 days):

  • Index (QQQ): (99 - 100)/100 = -1%

  • 3x Leveraged Index (TQQQ): (91 - 100)/100 = -9%

Share

Leverage vs. Concentration

Investors prefer concentration over leverage when maximizing their expected return.

This means most of us will allocate larger and larger portions of our portfolios to stocks, expecting higher returns. In contrast, you would have gotten a better risk-adjusted return if you had a well-diversified portfolio and then leveraged that.

One key reason for increasing concentration rather than leverage was aversion to leverage. Investors believed that increasing concentration was the “natural or conventional” path to taking leverage, which felt speculative.

What’s the optimal leverage?

While there is no one right answer, leveraging 10x on the market rarely works out. That’s because if you are levered 10x and the market drops 10% (relatively common), you will be wiped out.

In this thought experiment by Elm Wealth, the impact of using leverage in the S&P 500 from 1927 to 2020 was evaluated.

Notice that 1.5x leverage leaves you with the most money at the end, but at some point, you’d have experienced a drawdown of 95%.

Above 1.5x leverage, return goes down and risk goes up, driven by the volatility drag we’ve been discussing.

At 4x leverage, after suffering more than a 99.99% drawdown, you’d eventually only be left with the $1 you started with (just $0.09 inflation-adjusted), and at 5x leverage you’d have been fully wiped out on October 19th, 1987.

Most studies recommend up to 2x leverage, and there are few good reasons to go above 3x.

TQQQ during the Dot-com bubble and the Global Financial Crisis

The biggest concern regarding leveraged ETFs is that most of them were not present during the dot-com bubble and the Global Financial Crisis.

For example, TQQQ was launched in 2010.

So what happens if we create a simulation to evaluate how a hypothetical TQQQ portfolio would have performed?

The 3x-leveraged QQQ would have been wiped out twice — once during the dot-com bubble crash and again during the global financial crisis.

Leave a comment

$100 invested in TQQQ at the end of 1999 would have grown ~10x by the top of the dot-com bubble in 2001. But in the crash that followed, the portfolio would have dropped to <$1. While in our simulation, the portfolio does make a comeback, we would argue that, given the transaction and borrowing costs (which we ignored here), a real fund would have been shut down.

This also corroborates our earlier research, which recommends that optimal leverage be <2. Let’s take another example of a real leveraged ETF — SSO (2x leveraged S&P 500), which was launched in 2006 (before the Global Financial Crisis).

$100 invested in SSO at its launch would drop to ~$20 during the Global Financial Crisis. While this drawdown was brutal, it was not a scenario so extreme that the fund had to shut down. During the bull market that followed, the SSO portfolio overtook the S&P 500 by 2014, and by 2025, your investment in SSO would be 2x that of the S&P 500.

Researchers at Vanderbilt University had similar findings.

When they ran 10,000 Monte Carlo Simulations on leveraged ETFs based on the S&P 500, they found that over a 20-year period, a 2x-levered ETF had only a 0.5% probability of shutting down, and a 3x-levered ETF had a 7.6% probability of shutting down.

It’s also interesting to note that if you were 2x short the S&P 500, you had a 94% chance of the fund shutting down over a 20-year investment horizon.

Should you be investing in leveraged ETFs now?

Not if you are looking at individual stocks and especially semiconductors. Not to be “that guy”, but our MS portfolio has nearly doubled YTD without touching leverage.

There is no reason to add additional risk by using 2x or 3x leverage on an already volatile name. That’s how you end up in situations like this:

Finally, you have to be aware that for leveraged ETFs to make money, it’s not enough for the market to go up; it should do so with relatively low volatility.

Volatility Drag

While most investors are worried about the fund shutting down, the real risks with leveraged ETFs lie in the volatility drag. Take our earlier example of the fund's +10% and -10% moves over 2 days. The 3x levered ETF declined 9% compared to the underlying index's 1% drop.

In extreme cases, even if the index is moving up, you will lose money due to volatility. From February 2020 to January 2023, QQQ was up 30.76%. But, TQQQ declined 11.37%. This was due to

1. ~60% drop due to the COVID-19 crash

2. +200% run-up during the everything bubble in 2021

3. Another 50% drawdown during the 2022 rate hikes.

This incredible amount of volatility caused the 3x levered index to move very differently compared to the underlying index.

Leverage — The secret behind Buffett’s returns

Once again, it all comes down to your skill as an investor. While leveraging and investing in speculative stocks rarely works out, you can make leverage work for you if you know what you are doing.

While Buffett eschews debt publicly, a large chunk of his returns can be attributed to the smart use of leverage.

Berkshire Hathaway has realized a Sharpe ratio of 0.76, higher than any other stock or mutual fund with a history of more than 30 years.

In the famous paper Buffett’s Alpha, researchers analyzed Buffett’s portfolio to understand how he was able to generate such a high risk-adjusted return. They found that Buffett used the cheap leverage available to him from his insurance float and invested it in high-quality value stocks to generate market-beating returns.

We estimate that Buffett’s leverage is about 1.6-to-1 on average.

Buffett’s returns appear to be neither luck nor magic, but, rather, reward for the use of leverage combined with a focus on cheap, safe, quality stocks.

The real alpha might be in realizing that very few of us have the patience and the skill of Buffett.

Share

Market Sentiment is now fully reader-supported. A lot of work goes into these articles, and if you enjoyed this piece, PLEASE HIT THE LIKE BUTTON. It’s the only way new readers can discover our reports :)

Footnotes

2

Ignoring borrowing costs and transaction fees. Here is the exact mechanics of return compounding: (source)

• Quarterly revenue: $41.5B (up 340% yoy)

• Guidance for next Quarter: $50B (Estimate was $43B)

• 16 long-term contracts worth $100B+

Since the beginning of the year, we have consistently highlighted that memory has shifted from a cyclical product to a structural need for hyper-scalers, and Micron’s latest earnings prove it.

Why AI needs memory?

Reiterating our demand drivers from our memory bottleneck post,

• Agentic reasoning generates a long sequence of steps before giving the final answer. This means that the AI model should now retain all prior steps in memory so it has context for what it’s expected to do.

• Images, audio, and video generation consume far more memory than mere text generation. For example, to ensure characters are consistent throughout an AI-generated clip, the model must retain the entire frame in memory rather than just a few words.

• Finally, since model providers are pushing universal assistants that know everything about you, they now simultaneously have to keep the current conversation, general information, and old conversations to make sure the experience is seamless, which pushes the memory usage even further.

Nvidia’s own bill of materials tells the same story. The 2025 GB300 NVL72 rack carries about $374K of memory — roughly 9% of its ~$4M total cost. Just a year later, next-gen VR200 packs roughly 3x the memory capacity at roughly 2x the price per GB, pushing memory cost to ~$2M — a 435% jump that brings memory to ~26% of AI server cost.

New demand vectors

In the earnings call, Micron CEO Sanjay indicated memory demand for a robot is 10x that of an autonomous car. And these robots are expected to grow exponentially as the price per unit falls and they can tackle general-purpose tasks.

Powered by these tailwinds, we believe Memory will stay in a supercycle for years. That conviction is shared. Sixteen companies, including two hyperscalers, signed long-term agreements with Micron to lock in supply for three to five years. Traditional memory deals run two to three quarters at most — the shift reflects a long-term supply crunch.

These contracts cover advanced memory products and account for roughly 25% of Micron’s supply volume and 40% of its revenue at current prices. This is bullish for Micron as it secures prices and profit margins for the next three to five years.

The biggest issue with memory stocks was their cyclicality. Just 4 years ago, Micron’s gross margin went from the high 40s to near -31%, with Micron selling memory below cost due to oversupply. These long-term contracts prevent a similar downturn — The largest agreements carry a ceiling at the current CQ2 market price and a floor that holds through the term.

Even at the floor, Micron expects a gross margin well above its prior peak quarterly margin of roughly 61%, which puts the floor north of 65% on our estimate.

How are we positioning?

Read more

or in the “there is no rational reason for you to pay ~100x revenue multiple for a loss-making company.”

But what if there was a middle ground? What if you could trade this IPO without ever buying the trillion-dollar story? The setup is in the mechanics, not the fundamentals, and that's what we are going to find out today. But before we get there, let’s take a closer look at the company and how they make money.

Business Overview

While you can read the full 400-page prospectus here, SpaceX has 3 main business segments:

• Space — sells launch services. Revenue comes from two streams: (i) Launch Services, deploying payloads to orbit for commercial and government customers on the reusable Falcon 9 and Falcon Heavy, and (ii) Launch and Development, building spacecraft and providing launch/mission services for government space programs (Falcon 9, Falcon Heavy, Starship, and Dragon).

• Connectivity — is the Starlink business; revenue comes from broadband and Starlink Mobile service across 156 markets, split into Consumer Broadband, Enterprise Solutions, Government Solutions, and Mobile.

• AI — is the newly acquired xAI/X business. Revenue comes from advertising on the X platform plus AI solutions and infrastructure: subscriptions, data licensing arrangements (access to X's real-time and historical data), and API access to the Grok models.

Looking at the numbers,

• Starlink is the strongest business where revenue grew ~50% ($11B now), operating income more than doubled, and adjusted EBITDA rose 86%.

• Even though revenue improved for launch services, operating profit flipped from $21M profit to $657M loss — the driver here is Starship development, which increased R&D costs from $1.7B to $3B.

• AI revenue grew 22% (arguably small base of $2.6B), but now the operating loss is $6.3B. Once again, R&D is driving losses, as the company spent nearly $5B on training the Grok model and building out AI compute capacity.

This also allows us to build a rough valuation model to assign a value to each division.

SpaceX Model

Here’s how the numbers look:

• Starlink — We modeled subscribers to grow at a 25% CAGR from 9 million in 2025 to 84 million by 2035, a number that only holds if direct-to-cell scales alongside fixed broadband. Blended ARPU falls from $81 to $48 a month, because cheap direct-to-cell plans dilute the premium home base. At a blended 15x multiple, it carries $167 billion in intrinsic value.

• Launch Services — External Falcon 9 is anchored to third-party demand, near the 58 launches seen in 2025, rising into the late 2020s then declining as Starship absorbs the heavier external missions. Assuming a successful transition to Starship, modeled revenue reaches $25B by 2035, bringing the valuation for launch services close to $60B.

• xAI / AI — is the big gamble. As of now, the numbers are based on real contracts, but we are assuming a growth rate of nearly 50% for 10 years. 80% of the 2035 revenue comes from contracts that do not exist yet. Even with this optimistic projection, AI will add only $343B in value.

What’s interesting is that even with the most optimistic projections for each business segment, the company's total value comes to only $568B ($43 per share), a far cry from the asking price of $1.75 trillion ($135/share).

The kicker? If you exclude the AI division, the company's total value is around $220 billion — which is what it nearly traded at in private markets before the IPO hype and the xAI merger.

How to trade the world’s largest IPO?

While by valuation, SpaceX will be listed among the top-10 companies, it will trade like a mid-cap due to its small float. Broadcom, which is similar in size, has $1.7T in tradable float compared to only $75B for SpaceX.

Now that we know that the stock is going to be volatile and priced way above what it should be priced at by any reasonable estimates, here’s how you can trade it (based on whether you got the IPO allocation or not)

Read more

This is a guest post from . We launched Rebound Capital last June as our equity research division after more than 4 years of research at Market Sentiment. Here’s the one year update (and a special offer for Market Sentiment readers who make it till the end)

On June 5, 2025, we launched Rebound Capital. It was the culmination of more than 4 years of research at Market Sentiment, where we evaluated everything from investing in companies more than 100 years old, to economic moats, to tracking insider trades by members of Congress.

There’s this one idea we always came back to: investing in quality companies in a drawdown.

Here’s what we wrote about Meta in 2022:

No matter how you look at the company, the financial metrics are incredible. 80% gross margin and 30% ROE with growing revenue. But the stock is down 55% in the last 1 year.

I am not pitching it as a great business. They certainly have some incredible challenges ahead of them in terms of privacy and attracting a young user base. No one knows directly the solutions for their hardest problems—that’s why they’re the hardest ones.

But writing off a company based on two bad quarterly reports is just too short-sighted.

The same theme kept recurring:

• Netflix lost 3/4th of its value because it did not grow subscribers one quarter.

• ASML lost ~1/2 its market cap on tariffs and export restrictions.

• Amazon was down 56% after its first quarterly loss since 2015.

So we finally pulled the trigger and launched Rebound Capital, putting our own capital on the line. We followed Charlie Munger’s advice of “take a simple idea and take it seriously” and spent the past year going through hundreds of stocks to find the optimal rebound setups.

Fast forward a year, and we are outperforming the S&P 500 while holding top 3 positions in relative value names such as Amazon, Constellation Software, and MercadoLibre. To put this in context, 79% of all active large-cap U.S. equity funds underperformed the S&P 500 last year.

While it’s easy to talk a good game when your portfolio is outperforming, sticking to the strategy when it’s underperforming is the hardest thing to do as a public facing equity research platform. Q1 was hard on the Rebound Portfolio, where we underperformed the market by 9%. We continued to share our portfolio performance publicly throughout the period of underperformance (here are our Feb, Mar, & April updates). But sticking to our strategy helped us double down on quality names that rebounded strongly after the March dip.

Not to name names, but there are a lot of “gurus” who have Excel spreadsheets where the portfolio never goes down, and they magically exit at the right time. For us, we have maintained the policy of informing you of all our trades 24 hours in advance, and honestly, our biggest success till now is that nobody has ever said anything bad about us.

Overall, we are now in a good position with our portfolio compounding faster than the index while trading at essentially the same valuation. The headline numbers (weighted average of our portfolio holdings):

• Revenue growth, next 3 years: ~18%

• EBIT growth, next 3 years: ~19%

• EV/EBIT: ~27x, in line with the S&P 500

Here is the unusual part. We own a basket whose EBIT is set to grow ~19% per year over the next three years, compared with <13% for the S&P 500, and we pay no premium on the multiple to do so.

Divide valuation by growth, and the portfolio sits at ~27/19 = ~1.4, versus 27/11 = ~2.5 for the index (lower is better). In short, our portfolio compounds faster than the market while costing about the same, which on a growth-adjusted basis makes it materially cheaper than the index it is built to beat.

Your support has been incredible, and it’s what makes this work feel meaningful. While most creators are hush-hush about how much they make, we are going to show you our real numbers. After all, this is the easy bit after sharing our portfolio performance :)

As of today, we have more than 21,000 investors who subscribe to us, of which 969 investors pay to read our research, which translates to revenue of $358K (ARR). The spike you see in December is when Michael Burry became a paid subscriber to Rebound Capital and endorsed our note on Birkenstock.

Consider this your nudge to put your own work out in public. You never know who’s reading.

Finally, of all the metrics at Rebound Capital, our retention rate is the one that we are most proud of. ~90% retention at 6 months is unheard of in the newsletter space and is the best testament to the quality of the research we produce and the value you derive from it.

Right now, the market is in love with AI. Money is pouring into a narrow group of names at prices that already reflect the next 5 years of growth. To be clear, we are bullish on what AI can do over the long run. But when an equipment vendor that has always been cyclical gets priced as if the cycle has ended, the reward for chasing it is small, and the cost of being wrong is high. With many indexes changing their decade-old rules to accommodate hyped-up IPOs like SpaceX, indexing is no longer the safe, neutral choice. The case for doing your own research only gets stronger from here.

And this is where being independent matters. A fund manager has to answer to a committee, hug the benchmark, and own the hot name so they don’t look wrong at the next quarterly review. We answer to no one but you. We never have to chase the story of the moment and can sit in an out-of-favor name and wait for the market to catch up, because no one is going to pull our capital for being early.

While investors flock to one part of the market, high-quality companies are being ignored. Some of the best compounders we know are at their lowest multiples in years. We will continue to focus here. The crowd overpays for what it loves and ignores the good businesses on sale. But the market eventually catches on.

So we will keep researching the names nobody wants to talk about. The rotation back to high-quality names will come. Our job is to be patient enough to hold the right names when it arrives.

Anniversary Offer from Rebound Capital

To celebrate one year of Rebound Capital, we are offering 33% off to all current subscribers of Market Sentiment. This offer is valid only for the next 24 hours and if you are an existing reader of Market Sentiment.

Subscribe to Rebound Capital

With how the markets are right now, we are right in the middle of the good old days. In case you missed the everything bubble of 2021, this is right about how it felt back then.

Every thesis was proven in days, and random companies were doubling up every week. Hell, we were even trading pictures of rocks at the same value as that of an actual waterfront property in Florida.

While we are not that crazy yet (as the price rise is somewhat supported by earnings improvement), when we did our last portfolio update a month back, the MS portfolio was up 33% YTD.

Today, we are at ~78%!

The problem is that even though demand is real, there's a lot of crowding into semiconductor stocks. "Long Semis" is now the most crowded trade, and investors are already discounting multiple years of strong growth.

As Rebound Capital highlighted, there is literally panic buying in semiconductor stocks.

AMD is up roughly 330% over the trailing 12 months and now trades at 50x forward earnings, compared with Nvidia at 24x. The semicap names have rerated even harder: ASML at roughly 52x trailing earnings versus 38x at the end of 2024, Lam Research at 60x against a ten-year average of 20x, and Applied Materials at 42x against a ten-year average of 18x. These are cyclical equipment vendors being valued as if the cycle no longer exists.

The memory and optics complex is where the melt-up is cleanest. SK Hynix is up about 248% year to date and Micron 226%, with both joining Samsung in the trillion-dollar club. In optics, Applied Optoelectronics is up roughly 335% YTD, Lumentum 120%, and Coherent close to 80%. Some of the other Tier 2/3 suppliers are up similarly.

With the bad news out of the way, here’s what’s happening with the fundamentals:

Demand for semiconductors is nothing like what we’ve seen before. Until now, investors expected that we were at peak demand, with hyperscalers reinvesting almost all of their cash flow into AI data centers. With Google raising $80B from stock sales to fund its AI buildout, there is now theoretically no upper limit on the money that can be pumped into the ecosystem.

Still, even with the parabolic increase in revenue and profits, we're slightly jittery as the stock price tends to drop well before the slowdown data comes through in earnings reports. Based on the previous cycles, by the time the estimate cuts arrived, IT stocks had already fallen an average of nearly 30% — and individual names like Intel, Nvidia, and Micron fell considerably more.

With this level of crowding and hype, any hiccup, however temporary, will lead to a sharp sell-off. The optimal strategy here would be to take a step back, dig into names and sectors outside the current hype cycle, and invest at more palatable valuations.

Read more

Bonds might just be that.

So, with the AI trade in full swing and investors more than willing to pay, check notes, 100x revenue multiple on a loss-making company, it’s important to take a step back and evaluate what will happen if (when?) the Fed starts raising interest rates.

Interest rates are important once more

What most investors conveniently forget in a hype cycle is that a stock’s value is ultimately what the company will earn in the future, discounted back to today. So, when you are buying a company with a P/E of 50, it means that, based on current earnings, it will take 50 years for the company to “pay back” your investment.

You could argue that the 50 P/E company is growing rapidly, and your payout will be much faster. But the fact remains that most of your return is coming from projected cash flows that are 10+ years out.

Coming back to the interest rate bit, the moment rates rise, future earnings are worth less in today’s dollars. A company whose value is predominantly based on future profits will immediately get re-rated.

That’s what happened in 2022 once the Fed started raising rates:

In many cases, the company fundamentals actually improved — Shopify nearly doubled its revenue from $2.9B in 2020 to $5.6B in 2022, yet the stock fell ~70% because of multiple compression. In 2020, the company traded at a 50x revenue multiple, compared with an 8x multiple in 2022.

Same company, different interest rate environment.

Now look at where Semiconductor stocks are trading now:

We are in the best market for semiconductor stocks in the last few decades, and that has driven valuations too far for some companies. So even if the growth projections actually come through, if the Fed starts raising rates, all these companies will get re-rated.

What are the odds of a rate hike?

According to the CME FedWatch Tool, there’s a 70% chance of the Fed raising the federal funds rate by the end of the year. The heaviest odds (more than 40%) are on the Fed doing one quarter-point rate hike from the current target of 3.50%-3.75%. The good part is that there is only a 2% chance of a rate hike in the next FOMC meeting (Mid June).

The real worry is not the 25bps hike: it’s a combination of a new hawkish Fed chair, sticky inflation (due to tariffs and Iran war), and rising national debt — all hitting simultaneously when the semiconductor names are priced for perfection.

Read more

1. We are still early, and there is a long way to go.

2. We are near the top, on the cusp of a dot-com style collapse.

While no one can be certain, it’s probably somewhere in the middle. But if you are in the it’s a bubble bucket, let me start by highlighting the mistake I made at the end of 2023 when I published “The Problem with the Magnificent 7”.

Nvidia had ended the year with a 245% return and became one of the top 5 most valuable companies in the world. Nvidia’s Price/Sales had doubled from ~11x in 2022 to 22x in 2023. We were naturally skeptical of how much more the company could grow from these rich valuations.

Turns out, we underestimated just how massive the AI buildout would be. While we were busy looking at Nvidia's P/S and thinking it was expensive, the company's actual sales were about to explode. Revenue jumped from $61 billion in FY2024 to $130 billion in FY2025 and $216 billion in FY2026. Every big tech company was suddenly in an arms race to buy as many GPUs as they could get their hands on, and Nvidia was the only one selling. So even though the stock looked expensive at the end of 2023, earnings grew so fast that we missed the rally.

It took me quite a while to change my mind about AI and to understand that it’s a structural, not a cyclical, change. It’s often said that the mark of a good investor is the ability to change his mind when presented with new information.

So, before we get into the Agentic stack, here are 3 bits of information on how much AI is going to change the world in the next few years:

Market Sentiment dives deep into the bottlenecks of the AI world. Join 45,500 other investors to make sure you don’t miss our next briefing.

Subscribe now

1. Outside of the AI hype, we are quite a bit early

As per the latest estimates, ~78% of humanity has never used any AI tool. So while it feels like everyone is running an agent on tech Twitter, reality is far different.

2. Even with this, foundational company revenues are exploding

3. We haven’t even considered physical agents

Figure just had a 24-hour live stream featuring an autonomous humanoid robot picking and sorting packages.

Agents are going mainstream

Most AI conversations right now are about coding assistants (and that is what is driving the revenue surge for OpenAI and Anthropic), but that’s really only useful for the ~25 million developers out there.

Agents are going after something much bigger. The 1.4 billion knowledge workers: who spend their days buried in email, spreadsheets, and jumping between a dozen different tools. And once agents start handling even a slice of that work, AI usage goes through the roof, which means a lot more compute, a lot more GPUs, and a lot more demand flowing right back to the semiconductor stack.

On May 7, Anthropic announced that Claude is now generally available inside Excel, PowerPoint, and Word, with Outlook in public beta.

Up until a few months ago, using AI in your day job meant writing prompts, hitting an API, or running tools like Cursor. With this release, AI will now operate the 3 apps that 1.4 billion knowledge workers use the most. Microsoft has its own version of this, called Copilot. Copilot has ~20 million paying users, and they rolled out their latest Copilot and Office 365 E7 bundle on May 1st.

Watch the demo to appreciate what this looks like in practice. This is an inflection point for AI usage in everyday tasks.

Agentic AI: the inflection point

A developer running a focused 30-minute coding session on Cursor uses around 50K to 200K tokens (unit of compute). Compare that to a financial analyst asking Claude to reconcile a 200-row file against a pivot table, draft the email summary, and prep three slides for Monday’s review.

That single workflow uses 800K to 2M tokens.

Now multiply that by 1.4 billion office workers. Microsoft has already said that over ~80% of Fortune 500 companies have rolled out Copilot to part of their workforce. The load on enterprise systems caused by Agentic AI will make chatbots' token consumption seem trivial.

And the infrastructure to handle all this simply isn’t there yet. A chatbot session is quick, GPU-heavy, and low-context. An agentic session is the opposite. It runs for hours, constantly does things, and pulls on CPUs (to orchestrate actions), memory (for more context), and GPUs more evenly. As we wrote in ‘Notes on Compute’ last week, the server CPU to GPU ratio has already shifted from 1:12 in old training setups to 1:2 in agentic inference builds.

An explosion in token consumption

Goldman Sachs Global Investment Research now models that consumer and enterprise agents will push monthly token consumption to roughly 24× current global capacity by 2030. The crucial line on their chart is the dashed marker labeled ‘Token Economics Turn Positive in 1H26’, the moment any operator can deploy an agent whose output value exceeds its compute cost. That is when every company in the world will need to implement AI agents to compete.

Right now, agentic use cases are still more hobbyist tinkering than full-fledged commercial deployment.

Once an enterprise agent can operate continuously, 24/7, handling complex tasks and recoup its costs, adoption accelerates rapidly, probably within months rather than years.

Mapping the stack

Most readers already have meaningful exposure to popular names such as Nvidia, AMD, Intel, TSMC, Samsung, and SK Hynix, either directly or through our Advanced Packaging basket.

The Agentic Basket below is deliberately tilted toward the under-covered stocks, enabling agentic AI. These are the stocks with a higher chance of re-rating, as the broader street discovers them.

Morgan Stanley’s exposure map is the best reference we’ve seen for the full agentic stack. We’ll walk through each segment, then assemble the basket.

Read more

We are publishing this without a paywall.

It’s impossible to miss what’s happening in the AI compute market now. While QQQ is up 13% YTD, AMD is up 96%, Arm Holdings is up 117%, and Intel, a company that most investors wrote off, is up 206%!

The performance has been so mind-bending that even the President is posting stock gains.

Jokes apart, there is a fundamental reason these companies are getting re-rated — AI, specifically AI agents.

What is an Agent?

An agent is simply a clever bit of software that acts autonomously to achieve a certain goal. For example, if you want to find a house to rent in New York, you can spin up an agent and share your requirements (rooms, space, neighborhood, budget, etc.). — The agent will then browse the web, filter the houses based on your criteria, and give you a shortlist.

Before AI, automating this would have taken a few days of coding, and you would most likely just do it yourself. But with AI, you can just give the task, and it will write the underlying code and handle it for you.

Now imagine the potential if it could automate almost anything you ask it to. That’s what Manus promised to be (it’s worth watching the 4-minute video to get an idea of what an agent is and what’s possible). The demand for the product was so much that they had a 2M waitlist a few days after launch.

So why are compute stocks rallying?

Nvidia has more than 10’xd since the launch of ChatGPT. This is expected as we were using massive arrays of Nvidia GPUs to train and serve these Large Language Models. But AI has advanced from a chatbot that answers questions to an operator that takes action.

Coming back to the above example of a house search, the agent has to open up a browser, navigate to a particular page, scrape the data, compare the listings, and repeat this hundreds of times.

Most of these tasks load the CPU more than the GPU. So the ratio of CPU work to GPU work is much higher in an agentic system than in a simple chatbot.

According to Morgan Stanley Research, earlier server builds used 1 CPU per 12 GPUs. With rising agentic workloads, the ratio is now closer to 1 CPU per 2 GPUs, which is driving the CPU bottleneck.

Share

Where are we in the cycle now?

This is where it gets interesting. While the current cycle is definitely looking stretched (and we took some profits this week), the long-term outlook is promising.

Right now, agentic use cases are more hobbyist tinkering with an idea than full-fledged commercial deployments. The moment when (if?) an enterprise agent, which is on 24x7 and can do complex operations, we see an inflection point in token usage. When an agent can run profitably (i.e., the value of the work it does exceeds its token usage), adoption will explode.

Putting this together, we have a good outcome for hyperscalers. With the cost to process each token falling, the price per token stabilizing, and demand exploding, hyperscalers will be able to dramatically improve their margins.

The biggest overhang on the hyperscaler (GOOG, AMZN, META) stocks is whether the hundreds of billions spent on AI data centers will provide adequate ROI. The moment they show margin improvement, the companies will get re-rated (also the reason you see analysts increasing their price targets after the stock pumps. New information is passed to the valuation model, resulting in a new target price. There is no alpha in following price targets without understanding the underlying reason for the rerating.)

Leave a comment

The CPU Trade

If we are purely looking at exposure to CPU, there are only 4 companies in the field who is having meaningful exposure:

• AMD — is the market leader in data center CPUs, and the gap is expected to widen further with the launch of 2nm Venice chips. AMD should get a better share of the incremental revenue driven by agentic demand for CPUs.

• Intel — is staging a comeback with Xeon 6 processors being adopted by Nvidia and Google as the CPU "head node" in AI infrastructure. Plus, being strategically critical to the US does not hurt.

• Nvidia — has launched its AI CPU, Vera, designed exclusively for AI agents.

• ARM — launched its first-ever data center processor — AGI CPU focused on handling agentic tasks. Rather than going the raw-power route, ARM is developing a more efficient processor to reduce token costs.

But we would argue that much of the upside here is already priced in. AMD, Intel, and ARM all have more than doubled YTD.

A better trade is to build an agentic basket with:

1. CPU Compute

2. High Bandwidth Memory & Storage

3. Foundry & Chip design

4. PCB, Substrate, and Controllers

5. Security & Payments

This way, rather than crowding into one overstretched sector, it gives you clean, diversified exposure across the full agentic stack. This is exactly what we are working on, and we will be publishing the full report next week.

Subscribe now

Disclaimer: This is not financial advice or a recommendation for any investment. The Content is for informational purposes only, you should not construe any such information or other material as legal, tax, investment, financial, or other advice.

Market Sentiment is now fully reader-supported. A lot of work goes into these articles, and if you enjoyed this piece, PLEASE HIT THE LIKE BUTTON. It’s the only way new readers can discover our reports :)

Why?

Big-Tech earnings are crushing it

Last week's earnings from Amazon, Microsoft, Meta, and Google once again validated the strength of the AI super-cycle. All of these companies stated that they are supply-constrained.

• Amazon — Revenue of $181.5B (+17% YoY), beat by 2.4% | AWS Revenue of $37.6B (+28% YoY), fastest growth in 15 quarters.

• Google — Revenue of $109.9B (+22% YoY), beat by 2.7% | Cloud Revenue of $20.0B (+63% YoY), beat by 8.8%.

• Microsoft — Revenue of $82.9B (+18% YoY) | Azure revenue growth of +40%, beat guidance of 37–38%.

• Meta — Revenue of $56.3B (+33% YoY), beat by 1.4% | Ad revenue +33% YoY.

It’s worth taking a step back and reading those numbers again! We already had sky-high expectations on incredibly high base numbers, and yet these companies delivered. For example, Google Cloud's expectation was ~50% YoY growth on a $12B base, and it delivered a 63% YoY growth.

The strong growth has led hyperscalers to double down on capex, which bodes well for our bottleneck stocks. MS Portfolio (launched Feb 1st) is now up 33% compared to 7% for QQQ and 4% for SPY.

Also, Intel handed us our first 2x after it reported arguably its best earnings in 2 decades.

While there are some supply constraints, Intel is an asymmetric bet on the US government’s influence in the chip race. — Market Sentiment Jan’26

With that, let’s dig into why compute is becoming the new AI bottleneck, how we are positioning for the next few months, and what the bear case is if oil stays up.

Market Sentiment delivers data-backed, actionable insights for long-term investors. Join 45,000 other investors to make sure you don’t miss our next briefing.

Subscribe now

CPUs are the new GPUs

There’s a reason AMD is up 61%, and Intel is up 153% YTD — AI Agents.

Agents can now plan, reason, and execute multi-step tasks without user input. Planning and coordination tasks rely heavily on CPUs and can account for up to 90% of the processing time in an agentic system.

According to Morgan Stanley Research, earlier server builds used 1 CPU per 12 GPUs. With rising agentic workloads, the ratio is now closer to 1 CPU per 2 GPUs, which is driving the CPU bottleneck.

But, it’s not just AMD and Intel that stand to benefit here:

Read more

But one downside is that you will miss on relatively smaller companies whose deals won’t be big enough to show up in either 10K or news. To overcome this, you must do bottom up research where you go one by one to identify the common supplier.

That’s exactly what we did to identify MKS Inc. MKS products are present in 85% of all semiconductor production tools worldwide. Nearly every step in the wafer fabrication process has an MKS component inside it and as far as we can tell, we are the first to cover this company in depth.

Market Sentiment delivers data-backed, actionable insights for long-term investors. Join 44,000 other investors to make sure you don’t miss our next briefing.

Subscribe now

What does MKS do?

When TSMC builds a chip at its 3-nanometre node, the wafer passes through hundreds of process steps. Etching, deposition, lithography, inspection, cleaning. Each step requires a tool, and those tools come from equipment companies you may have heard of: Applied Materials, Lam Research, KLA, Tokyo Electron, ASML. These are the companies that collectively dominate wafer fab equipment spending, an industry worth roughly $133B in 2025.

But here is the thing. Those tools don’t work without subsystems. An etching chamber needs to maintain exact vacuum pressure, or the plasma that carves nanometre-scale features into silicon will behave unpredictably. That vacuum pressure is measured by an MKS gauge. The reactive gases that feed the plasma need to be delivered at precisely controlled flow rates, measured in nanolitres per minute. Those flows are controlled by an MKS mass flow controller. The RF energy that ignites the plasma in the first place is delivered by an MKS power supply.

If any of these are off by fractions of a percent, the wafer is ruined. At 3 nanometres, the margin for error is effectively zero.

MKS sells these subsystems to the equipment makers, who integrate them into finished tools and sell those tools to fabs like TSMC, Samsung, and Intel. MKS doesn’t sell to the fabs directly (for the semiconductor business). It sells to the companies that sell to the fabs.

It is, quite literally, the picks and shovels supplier to the picks and shovels makers.

MKS’s moat is exceptionally strong

The qualification cycles for semiconductor subsystems are measured in years. An equipment maker like Applied Materials doesn’t switch vacuum gauge suppliers because someone quotes a lower price. The gauge has been validated through months of process qualification at the end customer’s fab. The data, the documentation, the process recipes all reference that specific component. Changing it means requalifying the entire process. Millions of dollars. Months of engineering time. And most importantly, potential yield risk.

This creates an extraordinary moat. Not because MKS has patents (though it does have many), but because switching costs are so high that once you’re designed in, you stay designed in for the life of the tool platform. Often decades!

And yet, MKS trades at roughly 26x forward earnings, the cheapest forward P/E in the entire semiconductor equipment peer group (median forward multiple of ~34x).

One of the reason for the discount is the $4.4 billion debt load from an acquisition that, as we’ll argue, may turn out to be the smartest deal by any player in the sector. The market is mispricing MKS due to the debt-load, and the asset that it bought is structurally undervalued.

A far less visible layer of the Semiconductor world

There are really only five semiconductor equipment companies that matter. Applied Materials, ASML, Lam Research, Tokyo Electron, and KLA. Between them, they account for the vast majority of global wafer fab equipment spending.

But none of them build all their tools from scratch.

Below them sits a layer that is far more fragmented and far less visible. MKS, Advanced Energy, VAT Group, Inficon, Edwards, Brooks Automation. Each one commands a narrow, deeply technical niche. Vacuum. Power delivery. Gas flow. Measurement. Motion control.

The reason no single company spans all of these is that the physics is genuinely different in each domain. Building a mass flow controller that can meter corrosive gases at nanolitre-per-minute rates is a completely different engineering problem from building an RF power supply that delivers megahertz-frequency energy to a plasma chamber. The metallurgy is different. The sensor physics is different. The failure modes are different.

The equipment OEMs figured this out decades ago. If Applied Materials tried to build its own vacuum gauges, its own RF generators, its own laser systems, and its own gas delivery modules, it would need to maintain world-class R&D programs in half a dozen unrelated physics domains simultaneously. That would slow it down and dilute its focus on what it actually does best, which is integrating all of these subsystems into a finished tool and qualifying that tool at the customer’s fab.

So the OEMs outsource. And the subsystem suppliers specialize. And the result is an industry structure where five companies sit on top, visible to everyone, and a web of specialists sits underneath, visible to almost nobody, each one indispensable.

MKS, in our view is the broadest of those specialists. And that breadth is what makes it different from every other name in the layer below.

The Atotech Deal

In 2022, MKS acquired Atotech, a German company, for $4.4 billion. This was not a small deal for a company of MKS’s size. MKS’s EV was around $10 billion at the time of announcement of this deal. It loaded the balance sheet with debt, about $4.5 billion of it as of early 2026, which is why the stock traded at a persistent discount to its semiconductor equipment peers for years after the deal closed.

But the strategic logic was, in our opinion, genuinely brilliant. To understand that, we’ll need to understand what Atotech really does and how it integrates seamlessly with MKS.

Atotech is the global leader in advanced electroplating chemistry for PCBs and semiconductor packaging. Making a PCB requires plating copper onto its layers. That plating uses specialized chemistry, specific chemical formulations that have been tested and qualified for each manufacturer’s production process. Atotech makes that chemistry.

The business model works like this: MKS (through Atotech) sells specialized plating equipment to PCB manufacturers. That equipment runs on MKS’s proprietary chemistry. Once a PCB maker buys the equipment and qualifies the process, they are essentially locked into buying MKS chemicals for years. Management calls this a “high attach rate.” The industry term is more descriptive: it’s a razor-and-blade model.

Once the equipment is installed and qualified, the customer has no practical alternative but to keep buying MKS’s chemistry for years. The equipment sale is a one-time event. The chemistry revenue recurs for as long as the production line operates.

This is a fundamentally different business from MKS’s legacy semiconductor subsystem operation. Semiconductor subsystems are cyclical. When chipmakers cut capex, subsystem orders drop. But chemistry gets used up during production and needs to be constantly replenished. As long as the PCB factory is running, it needs chemistry. Management estimates that roughly 40% of MKS’s revenue now comes from consumables and services, chemistry that gets consumed in production and ongoing service contracts. This provides a layer of recurring revenue stability that the old MKS never had, which has also started to reflect in how markets are viewing the company.

Why the company is mispriced?

Now let’s dig into what has kept MKS’s stock at a persistent discount to its equipment peers (and why MKS is a strong buy at current prices):

Read more

This is a classic example of anchoring bias. Since you saw the stock trade at $100 and it’s now at $150, you mentally anchor yourself to the first price, which makes the stock feel expensive. Conversely, a stock down 50% feels like a bargain because it's cheap vs. the anchor. What’s interesting is that none of these decisions is made by evaluating fundamentals, but rather by going purely by vibes.

Most of the companies we invest in at Market Sentiment trade near their all-time highs — and that's by design. We focus on businesses with bottleneck-like characteristics, where the market has yet to fully price in the long-term scope of their dominance.

And all this is not just theory: Here is our full portfolio holdings as of today. Instead of betting on which model is going to come out on top, it’s much more profitable to bet on the underlying companies that the AI boom runs on. The other thing worth noting: 15 out of 16 positions are in the green (while QQQ is down YTD), and there's no single name carrying the portfolio.

With that, let’s get into a few names that are instrumental for the AI buildout. We already have one of these in our portfolio, and we will add a few more in the coming weeks.

4. Corning Inc ($GLW)

Corning is a 175-year-old materials science company best known for Gorilla Glass — the scratch-resistant glass on virtually every smartphone screen. But recently, the company has quietly become one of the most essential infrastructure companies in the AI buildout. The reason is fiber. AI data centers require up to 10 times more fiber connectivity than traditional cloud computing, and Corning makes the glass it runs through.

The big break for the company came in January, when it announced a $6B deal with Meta, under which Corning will supply optical fiber, cable, and connectivity solutions for Meta’s AI data centers. Meta's 5-gigawatt Hyperion data center under construction in Richland Parish, Louisiana, alone will require 8 million miles of fiber.

While the Meta deal is already priced in, and the stock is definitely not cheap (PE of 92!), we expect a few other players, like Google and Amazon, to make similar deals. Another tailwind for the company is that newer, more powerful chips require more fiber-optic connections. For example, Nvidia's 72-GPU Blackwell nodes require 16 times as much fiber as traditional cloud switch racks. Finally, the geopolitical angle is also there, with Corning manufacturing optical fiber domestically, including two of the world's largest fiber facilities in North Carolina that employ over 5,000 people.

3. SK Hynix ($HY9H)

We already did a deep dive into SK Hynix 2 months ago, and it’s our largest portfolio position (up 22.4%).

The company primarily makes DRAM (Dynamic Random Access Memory) and NAND memory chips (long-term storage). Traditionally, most clients for these products were smartphone and PC manufacturers, and the market was chugging along just fine.

Then AI changed everything — models required enormous amounts of memory to operate, and AI-generated images and videos required significant storage. With companies spending hundreds of billions on capex, demand for DRAM and NAND skyrocketed, far outstripping supply.

As a result, SK Hynix’s revenue tripled from $22.8B in 2023 to $66.3B in 2025.

The thesis of our investment was based on SK Hynix's domination in high-bandwidth memory. With over 50% market share, SK Hynix has significant pricing power, enabling it to set terms and expand profit margins while maintaining its market share.

The memory bottleneck is only expected to worsen as:

• Agentic reasoning generates a long sequence of steps before giving the final answer. This means that the AI model should now retain all prior steps in memory so it has context for what it’s expected to do.

• Images, audio, and video generation consume far more memory than mere text generation.

• Finally, since model providers are pushing universal assistants that know everything about you, they now simultaneously have to keep the current conversation, general information, and old conversations.

FYI — SK Hynix is primarily listed on the KOSPI (Korea Composite Stock Price Index) under the ticker 000660. For international investors, the most accessible entry point is through Global Depository Receipts (GDRs). These are listed on the Frankfurt Stock Exchange under the ticker HY9H.F. They have filed for ADR listing in the U.S. market (Expected by mid 2026).

Micron is a decent alternative if you can’t access SK Hynix (for the Schwab investors). We have a small position there.

2. Hammond Power Solutions ($HMDPF)

Here’s a not-so-fun fact: Nearly half of planned US data centers are now being delayed or canceled due to shortages in critical electrical components like transformers, switchgear, and batteries.

Read more

Q1 2026 has not been kind to equities. Amid geopolitical uncertainties and rising oil prices, the S&P 500 is down ~4% YTD, and the Nasdaq is down 5%. We had briefly entered a correction territory for QQQ before the bounce back in the last two days.

Our bottleneck thesis continues to shine, as even with the drawdown, all three of our baskets have outperformed the market.

1. Advanced Packaging +4.5% vs -6.7% QQQ

2. Optics +0.7% vs -6.6% QQQ

3. Energy +1.5% vs 0.6% SPY

MS Portfolio Holdings Update

March was tough for the portfolio compared to Feb as geopolitical crisis muted AI investments and Google launched its TurboQuant AI-compression algorithm, which is expected to reduce the LLM memory usage — a headwind for our memory thesis. Even with these headwinds, our portfolio is in the green for the year while QQQ is down 5%.

With Oil price rising a whopping 78% this year and the crisis far from over, here’s how we are positioning our portfolio:

Read more

We are publishing this without a paywall.

This has now become an yearly ritual — exactly one year back, we had issued our buy the dip report when the S&P 500 was down 10% from it’s ATH as Trump had just started yet another trade war. Barring the global financial crisis, it was the worst start to a presidency in the last 50 years. Yet, one year later, the S&P 500 is up 13% (and 27% up if you had bought the liberation day dip).

We are once again looking at a similar playbook with Trump doing something irrational causing the market to drop. YTD, S&P 500 is down ~9% and QQQ is now down 12%.

One of our favorite thought experiments during times of extreme market uncertainty and fear is:

What trades do you expect to regret not making this week… a year from now?

Leave a comment

The key question in a market correction is whether to wait out the storm before investing again or double down and buy more?

To find out, let’s consider three investors: Cautious Charlie, Average Andy, and Daring Dave.

Each invests $100 in the S&P 500 at the beginning of every month. When the market falls below 10% of its previous all-time high (correction territory), each investor reacts differently to the dip.

• Cautious Charlie “holds” - He stops investing, holds the cash, and waits till the market crosses the threshold again to deploy the capital.

• Average Andy “stays” - He continues investing as usual.

• Daring Dave “doubles” - He invests double the usual amount till the market crosses the threshold again.

Who did better?

Here’s how they would have performed if they started investing in 2000:

At the end of 25 years, Dave, who doubled down during all the drawdowns, would have had the highest return, at 523%, compared to Andy’s 453%. Finally, cautious Charlie, who waited for the market to “stabilize” before deploying his capital, would have had the lowest return, at 295%.

This stark difference in performance is due to the underlying behavior of the stock market. According to JP Morgan, just missing the 10 best days in the stock market over the past 20 years would have dropped your return from 548% to 197%.

The kicker? Seven of the 10 best days took place in bear markets. Cautious Charlie, by trying to wait out the worst unknowingly missed out on some of the best days as well. Daring Dave, who bet hard when the markets were down, came out on top over the long run.

Baron Rothschild, an 18th-century British nobleman, is credited with saying, “Buy when there’s blood in the streets, even if the blood is your own.” He should know — He made a fortune buying during the panic that followed the Battle of Waterloo.

The worse things are in the market, the better your chances of finding a great deal. And things sure do look bleak now…

Disclaimer: This is not financial advice or a recommendation for any investment. The Content is for informational purposes only, you should not construe any such information or other material as legal, tax, investment, financial, or other advice.

Market Sentiment is now fully reader-supported. A lot of work goes into these articles, and if you enjoyed this piece, PLEASE HIT THE LIKE BUTTON. It’s the only way new readers can discover our reports :)

Share