The Hyperscalers

Let’s get this out of the way first. Google, Amazon, and Microsoft have been aggressive investors in frontier labs from the very beginning and have been adding along the way.

TL;DR: These stakes are a nice kicker to already great investments that also cause earnings surprises, but they are not a proxy. Maybe if Anthropic 5x’s one day.

Amazon: 20% stake, 5% of market cap

• ~$8b in since 2023, plus another $5b in early 2026 with a commitment for up to $20b more.

• Stake is 15-21% fully diluted but Amazon hasn’t disclosed the full figure

• Stake is worth $135-160b at the $965b mark, which is only 5–6% of Amazon’s ~$2.65T market cap

Alphabet: 15% stake, 3% of market cap

• ~$3b in + $40b more committed in April 2026 ($10b upfront at a ~$350b mark + $30b with milestones). Started with $300M back in early 2023.

• Stake: ~14% fully diluted, likely ~15% now

• Drove nearly half of Alphabet’s $62.6B Q1 2026 profit

• At the $965b valuation, stake is approx $135–145b. This is only ~3% of Alphabet’s ~$4T+ market cap

Single Stock Proxies

Now on to my favorite option. There are various companies that have made venture style investments in Anthropic. Let’s look at the top two.

SK Telecom: 0.5% stake, ~50% (?) of market cap

This is one of my favorite investment ideas over the past 6 months. was a telecom similar to the AT&T of Korea, mature and regulated. In recent years it has aggressively attempted to re-invent itself as an applied AI company and AI investor.

Here are some tidbits:

1. It has been through some things such as a major hack

2. The hack saga seems to be over and the company has started to recover financially (confounding variable)

3. It is part of an opaque family conglomerate structure known as a chaebol

4. We don’t know precisely how much of Anthropic is owned but estimates from analysts and the media are centered around 0.3-0.7%

5. Is subject to Korean currency risk

6. Is caught up in the speculative fury of the Korean retail market

7. Is subject to Strait of Hormuz impact as the country has a lot of energy exposure

8. There are all sorts of discounts applied to any embedded stake due to opacity, lack of clarity on shareholder returns, tax complications, reinvestment risk, etc.

9. There are other things going on with the company such as completed acquisitions and other programs to unlock value that confound pinning down the “look through”

All the uncertainties aside, the three hard parts are a) ascertaining their current fully diluted stake, b) guestimmating the relative allocation of market cap to this stake vs everything else now that other people have noticed, and c) forecasting where Anthropic will be after the IPO. Let’s take a closer look. First, the stake, which is disputed but a bit more objective.

Now what is this stake worth:

Thinking out loud: SKM was trading @ $20 for a $7.5b market cap before Anthropic went parabolic and now trades at about $40 or $15b market cap.. so if you SWAG $7.5b for the core business and 7.5b for the stake, 0.5% * 1.5 trillion = $7.5b, back of napkin checks out for this being a 50% (and growing) proxy for Anthropic. Presumably, if Anthropic makes it to $3T post-IPO you would see upside of 50% in SKM. This is just my mental math.

The biggest problem here is the circularity. Is the current market cap $7.5b core + $7.5b of recognized Anthropic, or is it $13b core (due to the idiosyncratic developments of the company) and $2.5b Anthropic? If it’s the former, then the company is a purer proxy so you get more direct delta exposure if you just want to be an investor at the current $1.5T whisper valuation. If it’s the latter, then there’s more “deep value” to be unlocked when the position is finally revealed after the IPO. I guess both are bullish for different reasons.

Example expression: debit spreads 35/55 in September or January.

Zoom: 0.33% stake, 18% (?) of market cap

Zoom is a little bit cleaner and simpler, but I was concerned the stake was too small.

Tidbits:

• Invested $51m in May 2023 @ $4.5b valuation

• Topped up ~$46m in early 2026

• The concrete fact that we need is the 10-Q which marks the stake at $1.3b against Anthropic’s $380b round. We can do the math directly: 0.33%

• The rest of it is clean.. mature biz growing 3% a year, $8b in cash, buybacks, etc.

• It was $84 last fall and $94 now… cherry picked to be before Anthropic went parabolic with Claude Code and after the Iran war recovery.

Similar circularity issue as before.. but if we assume investors already see the 0.33% and price it at $1.5T, that would be about $5b of a $27b market cap, or ~18%.

I think this just might be a bit too dilute to do the trick. Might be a good deep value play for someone but I want raw deltas.

Hyperliquid

One of the more interesting and direct ways is perpetual futures against tokens that track the market cap of pre-IPO companies.

The tokens themselves track an “Oracle” of private valuations that in turn are turned into a futures market. These particular futures provide 3x leverage on posted margin. There are a lot of complications:

1. Hyperliquid is not technically accessible to US investors and to try to access it is a violation of their terms of service. Any workarounds would be technical and carry their own risks.

2. I’ve heard these are crowded trades, so there is a high cost to carry them known as the “funding rate”

3. Possible liquidity issues, volatility, issues with the Oracle used, etc.

“Perhaps with just 3x leverage I could avoid being liquidated?”

“Perhaps I could set up an AWS instance and vibe code access to the DEX directly?”

Perhaps that’s all a bit too much.

P.S. Investing in the HYPE token itself, treasury companies that accumulate it ($PURR) and the related HYPE ETFs is a popular thing as well right now as folks are using Hyperliquid for price discovery of pre-IPO companies and things like oil during the Strait of Hormuz crisis when normal futures markets are closed.

Prediction Markets

One creative idea is to synthesize a long position by making a series of binary bets on relevant prediction markets.

This is likely not a serious idea due to lack of liquidity, access, and myriad other issues, but it’s a fun thought exercise.

What other creative ways have you seen and heard? Would love to hear from some skeptics as well.

—

The narrative of local, smaller, cheaper open source models as an alternative to overpaying for frontier tokens keeps popping up as a bear case. I think it’s a seductive and intuitive idea: why use a blowtorch to light a cigar? We have a focused job to do, so why all of this general intelligence, scale, cost, latency, and additional risk? Let’s have a small, affordable, lobotomized intelligence for each task.

For focused tasks like summarizing small chunks of text or general data munging, fine. But for coding and agentic workflows operating over longer horizons? I don’t think so.

I see three dynamics:

1. Total Cost to Resolution (TCR): for complex tasks, the more powerful model will likely ultimately cost less. Stronger models require fewer reasoning tokens, one shot the best solution more quickly, and reduce the number of brute force generate-verify iterations.

2. The Horizon Wall: Over long trajectories, these small differences have a big impact on quality and thus how long the horizons can be that the models work independently. Small errors compound into catastrophic impact on viability.

3. Parallelism Capacity: Because of B, smarter models allow more parallelism because they can work longer between human (or agent) review.

I think the compounding nature is the crucial point because it propagates over every phase. A poorer research session compounds into poorer planning which compounds into an inferior implementation and weaker code reviews and so on.

The math: if you compare a 95% success rate to an 80% success rate over a simple 10 step procedure, the former will succeed 60% of the time and the latter 10%. What if the chain is hundreds or thousands of steps?

Perhaps harnesses will become more model heterogeneous (likely at the behest of the orchestrating models themselves) to reduce latency and cost. For planning and agentic trajectories, there is no substitute for the best.

TL;DR: The idea: 50bps of equity in premium: December 2028 800C. You can call it a lotto ticket, I call it underpriced right-tail skew. Yes, that’s a 10T market cap, but the position can be monetized long before.

Backstory:

I was really surprised to see this company in the mid 350s again. Perhaps some of these high quality megacaps were already being sold in anticipation of the major IPOs or to fund the AI capex rally.

Alas, there was an $80b equity offering announced yesterday. Quick facts:

• $10b @ 350 by Berkshire Hathaway - note they are adding to their earlier $4.3b then $11b investments in 2025 and 2026. They are averaging UP. Greg Abel is coming out of the gate strong with a sure bet here.

• $30b underwritten

• $40b at the market over time ($30b earmarked for employee tax obligations)

Management must feel the same way about the leverage they will see from this additional capex to take this step. Beyond simply buying the dip, I think the real opportunity is further out.

If you look at the 2028 out-of-the-money LEAPS, their IV is about 38%. When compared to historical vol at 35% or less, you could argue that this is rich. I don’t think so.

For the 600C, it would require modest multiple expansion and hitting the consensus bull case on EPS but nothing dramatic to break even. These contracts only price a 1 in 5 chance that they hit $600+.

However, since this opportunity crosses a crucial phase in the development of this theme, there could be a lot of right tail risk up for grabs… such as:

• TPUs revealed as a big competitor to NVIDIA

• Search continues to re-accelerate

• Google cloud continues to go parabolic

• Waymo cracks self driving

• Gemini 4 tops Mythos

• .. and so on

There is more upside skew to this company beyond projecting current financials forward.

By going all the way up and all the way out on the chain, there’s a chance to capture the direction of the base case as well as be long vega if Google rerates to an IV of 45-50% like Nvidia/AVGO etc. Regardless of spot price, these could gain 50% on IV alone.

650C is probably the more reasonable choice, but 800C is offering the maximum capital efficiency, omega, notional, and vega exposure dollar-for-dollar. Sized accordingly, this gets me over the hump of coming back to a concentrated position.

In brief, we are waiting for a moment of upside panic. A surge in spot and an expansion in long term IV.

Appendix, more cool charts:

In the past week many are choosing software for this rotation. I view this as highly questionable due to the extreme cone of uncertainty, although I do hold a few potential agentic winners with operational and market momentum such as , , . If there are other winners, I am happy to let you take the first leg up and scale into the success later. Open to ideas.

Let’s take a look at the menu:

• Frontier labs: not easily investable, my preferred vehicle is which has become an Anthropic equity holdco. Might also be a Hormuz peace dividend in there too.

• Picking software winners: as mentioned above, trying to pick through the pile and ID the cleanest sure bets and avoiding the risk.

• Gold miners such as or - these were really beaten up by Hormuz so many have a lot of upside upon resolution. The fundamental case has only been strengthened by the conflict.

• Silver: as the switch-hitter narrative of energy demand, debasement, and retail frenzy with a similar Hormuz recovery story waiting in the wings. Straddles the AI trade and the debasement trade more directly than Gold

• Neoclouds: they are pure plays to the inference demand which is repeatedly re-affirmed. I expected these to be the most convex and successful bet in the book but alas they have chopped around and been met with sellers on earnings beats etc. If the wall of worry fades around their economics, debt, etc. there may be a lot of energy here. has started to show signs of breaking out. I am already heavy here, so alas. has been on a roll as well, with a large 15% position shown in the Situational Awareness LP 13F.

• Nuclear: such as - a much bigger secular story than the AI trade alone, especially in the wake of Hormuz.

• Hyperscalers: consistently crushing earnings reports and sitting in the nexus of supply, demand, and application. Some of these are on sale such as MSFT (I am equal parts intrigued and concerned), some have run hard but pulled back and appear to have a 2x on the horizon such as Alphabet (personal top pick). The IV for longer dated options is also underpriced IMO considering their positioning. Through LEAPS, I play IV/RV and direction together.

• Energy IPP like and - lacking YTD momentum, so a possible way to reduce the momentum factor.

• Leaders in orthogonal sectors like Healthcare / Biotech: I love as a natural pairing. Low correlation, ignores semis pullbacks.

Potential thesis: all of the yellow bars might be a bit better than they look.

Through these historical periods:

Biotech, infra, and energy look OK.

And a little more color:

Just food for thought.

Sustaining this exposure can sometimes feel like momentum chasing or meme stock mania. I have to replenish my conviction from time to time to remember what the Alpha is, which is front-running the mid to long term reaction to what is unfolding.

What we already know:

A few current events

• Google just reaffirmed their expanded Capex plans, cloud backlog (2x due to Anthropic later widely cited) and indicated an unspecified INCREASE for next year: Q1 2026 GOOGL call (April 29, 2026)

  Anat Ashkenazi, CFO: “We are updating our full year 2026 CapEx guidance range to $180 billion to $190 billion, up from our previous estimate of $175 billion to $185 billion to now include investment related to the acquisition of Intersect…”

  “We expect our 2027 CapEx to significantly increase compared to 2026.”

• NVIDA re-confirmed the same, forecasting 3T in total annual Capex by 2030 (not just hyperscalers), FY27 Q1 (May 20, 2026):

  “With analysts now forecasting hyperscale CapEx to exceed $1 trillion by 2027 and Agentic AI beginning to proliferate all industries, AI infrastructure spending is on track to reach $3 trillion to $4 trillion annually by the end of this decade.”

For the risk tolerant, these breadcrumbs justify ratcheting up the bets on AI infrastructure, semis, and hyperscalers over the next few years.

But is it inelastic?

Perhaps we should take the time to look at what’s happening that may preclude any of these major players from pulling back in the foreseeable future regardless of the macro regime.

1. Claude Mythos

Beyond the cyber-security aspects and Project Glasswing, look at the benchmarks:

That’s a gigantic gap up - regardless of nuance around cost to serve. Not just holding, but a discontinuity. Google, who appeared to be in the drivers seat with Gemini 3, now feels 6 months behind. The race is on.

Look at this chart from METR (if you aren’t familiar with METR trendlines, brush up):

Does this look like a good time to cancel planned capacity you can afford in any case and with cloud revenues growing 63% YoY?

2. An OpenAI model has disproved a central conjecture in discrete geometry

We had IMO gold, now this.

One of the confirmation signals I have been personally waiting for is a unique contribution that isn’t long tail:

• an autonomous company turning a profit

• novel drug discovery

• cracking of a major unsolved proof like a millennium problem

• Materials science, even if shy of a room temperature superconductor

TLDR; Paul Erdős was a prolific mathematician that posed over a thousand problems (half unsolved) that have attracted various levels of professional human effort. With (disputed) buzz around models tackling some of these, OpenAI disproved a major one.

While this feels halfway there, there’s an asymmetry for me. It seems to tip the scale from “possible” to “nearly guaranteed” re: the revelation of a major discovery or solution within the next few years. These headlines will likely continue and their implications will sink in.

UPDATE: Google DeepMind just announced they have paired LLMs with the Lean language to solve 9 more, including 2 that have been unsolved for 56 years. Not an expert, but the headlines are trickling in.

What does this have to do with capex being inelastic?

The decision makers in this race see these trendlines and capability discontinuities and can’t afford to be caught out. It’s existential. This isn’t a new argument, but it’s probably essential to the idea to doubling down over and over on a cyclical sector.

A counterargument in the mid-term

What about the marginal AI infrastructure Capex customer? It’s pretty clear that the risk/reward for all of the major players will result in full steam ahead regardless of the macro picture, but what about the marginal buyers if a bond scare or headline scare like the Saaspocalypse triggers some reflexivity? In the mid term, I think it is worth defining who those marginal customers are and watching closely. Stay tuned.

I was long NVDA again for 3 hours yesterday. The company is of the highest caliber, has not participated as much as they deserve in the fantastic rally in semis this season, and seemed ripe for some recognition. The thesis was to capture any information reveal then move back to more convex opportunities.

Numbers out, crushed without a blemish, reaction muted. Moving on.

Market cap is one of my favorite vibe metrics for eyeballing an opportunity. In fact, my biggest successes this year (AMD, Intel) were due entirely to this phenomena. If you project the happy path forward where NVIDIA is a 10 trillion company, does it make sense that Intel (the only credible advanced domestic foundry) and AMD would hold a market cap of 100b and 250b respectively? No. One way or the other they would chip off more than 3% on a relative basis.

Don’t settle

If I can’t tell a simple story as to why a play has a chance of being high-skew over the next 2 years in addition to clear momentum and thematic fit, then I’ve given up before even starting.

Marvell and Nokia seem to be two candidates that have the right headroom, added both. Going to circle back on that later but found good primers out there from if you’re curious why.

—

Full microsite here: https://macrodesk.ai/share/nvda-too-big/dacbba12/c

—

Subscribe now

Allow me to give a concrete example.

1. AVGO is really interesting, Mythos was trained entirely without NVIDIA in the stack. Let’s lever up with some 50 or 35 delta LEAPs.

2. Intel is a no brainer, it’s the only credible path to domestic advanced foundry. Add LEAPs.

3. NVDA is cheap right now, it’s being punished for its size and stickiness to the S&P, in the short to medium term it certainly has room to run. More options.

4. Memory isn’t cyclical, it’s a new secular trend trading at 8x fwd P/E. Let’s ride the momentum and pick up some shorter tenor debit spreads.

5. AMD is a catch up trade, whatever it’s worth, if NVDA is going to 6 trillion then AMD must go to 1 trillion, 100% upside.

6. .. and so on.

Next thing you know you’re holding 15 names… if these were normal equity positions that would be no problem, and perhaps desirable for tax loss harvesting. However, if you are using options to lever up your theme beta, you could be overpaying dramatically.

2. What’s the issue?

When you place a bet in this way, you’re making a directional bet but also betting that the magnitude of the move over that period is larger than what the market expects and has priced. Take SOXX as an example, which has a generally even 5-10% allocation to the above ideas.

• SOXX may have a 43% IV for LEAPs 2 years out

• Micron on the other hand may have an IV of 80%+ over the same period, same for Intel AMD etc.

If you were to bet on every name in this way on a pro-rata basis, you may be paying an average IV of 64% vs. 45% for the ETF itself. People pay a huge premium for the privilege of betting on single name outperformance.

3. Why does this happen?

Because individual moves partially “cancel out” inside of the ETF, the ETF itself is less volatile and carries lower premiums. There are professional “dispersion traders” that specialize on isolating this and trading it directly but we just need to be aware of it.

4. How I approach this

When I am bullish on a dozen names in the ETF for distinct reasons, this is actually the ideal time to bet on the ETF itself. Because the IV implies winners and losers when I see mostly winners (i.e. correlated), there’s a discount to harvest. For the single name bets, I limit these to the 2 or 3 that I feel will not only beat but CRUSH their individual IVs. Generally this means multibag potential for something like a Micron or Intel. I had that conviction on Intel when it was trading below a 100b market cap, but now I’m more uniformly bullish on it vis-a-vis the rest of the names in SOXX after a 450% rally.

5. The key point

You need to be VERY bullish on your single names when you’re paying up for high-IV names. Otherwise you may be better off betting on the sector itself. Don’t open more than a few positions until you understand this.

6. Bonus: what’s worth the premium to me today?

• I like Marvell due to it being a mixture of the Intel and AVGO thesis - has the market cap shape to multibag and is in place for the explosion of alternative ASICs. Worth the extra premia.

• I also like AVGO for the opposite reason, IV is actually looking very attractive (comparable to SOXX itself) but with a higher level of conviction vs. the sector itself.

• I feel the same about Oracle but that’s a story for another day. Perhaps could chime in on this one.

What companies are you willing to pay 100% IV for?