by Top team

Thiel bet his portfolio on the outlet: chips and electricity

Peter Thiel filed a $418.7 million quarterly report that holds no AI stock, no chipmaker and no cryptocurrency. Seven of eight positions are power utilities. One quarter earlier, the same filing held nothing at all.

On August 14, 2026, Peter Thiel's investment vehicle, Thiel Macro LLC, filed a quarterly holdings report with the US Securities and Exchange Commission, Form 13F. It lists eight lines totaling $418,667,000. Seven of the eight are power utilities plus one small modular nuclear reactor developer. Not one chipmaker. Not one AI company outright. Not one cryptocurrency, despite Thiel being known for years as one of bitcoin's earliest institutional backers.

What wasn't in the report one quarter earlier is even more telling. We pulled the fund's two previous filings, for the first quarter of 2026 and the fourth quarter of 2025. Both hold an empty table: a single NA line, zero value. That doesn't mean Thiel held no other assets at all; 13F doesn't capture private deals routed through Founders Fund and other vehicles. But the specific reporting line we can now see in full was built from zero in one quarter, and it landed entirely in a single sector.

One investor doesn't make a market. But this particular investor, a Palantir co-owner, a PayPal co-founder and an early Facebook investor, put $418.7 million into a corner of the market where three independent trails of evidence converge on the same fact: bitcoin miners' hyperscaler contracts, memory and chip-packaging producers' own numbers, and data centers' own complaints about grid queues. The limit on the AI industry in 2026 isn't the chip. It's the outlet.

How we counted

Every fund number here, Thiel's and the five others, comes straight from their own SEC filings for Q2 2026, not from a third-party dashboard. The numbers on miners, memory and chips come from company press releases and industry trackers (CoinShares, TSMC, SK Hynix); wherever a figure is an estimate rather than a confirmed number, the text says so.

Thiel's portfolio

Here are all eight positions, ranked by share of the portfolio.

Add up the five utility names in the middle of the table, AEP, DTE, FirstEnergy, CMS Energy and Vistra, and you get 52.8 percent of the portfolio parked, literally, in the wires and generators of the American grid. Add Vista Energy, another 18.1 percent in the raw material that feeds that grid, and almost three-quarters of the portfolio turns into a bet that electricity stays scarce, and that whoever's willing to pay for it will have to keep paying.

Amazon looks like the outlier on this list: the largest single position and the only recognizable tech name. But by the company's own 2026 disclosures, AWS is the world's largest corporate buyer of electricity for new construction, and the cloud is competing for the same megawatts as the other seven positions in the portfolio. Thiel didn't buy AI. He bought the thing AI physically cannot run without, and he bought it seven different ways at once.

X-Energy is the smallest position, under one percent of the portfolio, but it carries the longest time horizon. The company is developing the Xe-100, a small modular reactor whose build cycle, unlike a conventional nuclear plant, is measured in years rather than a decade. In January 2026, Meta announced a similar bet in public: three nuclear power agreements totaling more than 6 gigawatts, including a twenty-year deal with that same Vistra Corp for 2.1 gigawatts of capacity from plants already in operation. Thiel didn't buy the idea of a future reactor. He bought the company building it, and the company already selling megawatts from working plants to whoever will sign a twenty-year contract.

Broken down by asset type, the portfolio reads even more plainly: 52.8 percent is wires and generators (AEP, DTE, FirstEnergy, CMS Energy, Vistra), 28.2 percent is the cloud and its appetite for power (Amazon), 18.1 percent is the raw material for generation (Vista Energy), and 0.9 percent is tomorrow's generation itself (X-Energy). Not one percent went to the chip itself.

Who else holds the same names

One fund's portfolio could be a coincidence. To check whether it is, we pulled the same Q2 2026 13F tables from five more managers with reputations that couldn't be more different from Thiel's: Soros Fund Management, George Soros's fund, an investor Thiel has publicly called his ideological opposite; Stanley Druckenmiller's Duquesne Family Office; Renaissance Technologies, the world's largest quant fund; Cathie Wood's ARK Investment Management; and Coatue Management. The method was identical: not a rehash of somebody else's dashboard, but our own read of their SEC filings.

The result: five of Thiel Macro's eight holdings also turned up in Soros's filing for the same quarter.

*Source: 13F filings for Q2 2026, pulled directly from SEC EDGAR. Share is each row's value against the fund's total disclosed portfolio value on the same date. "Not found" means the stock doesn't appear in the fund's disclosed holdings.*

A five-out-of-eight overlap with a fund whose public reputation runs the other way isn't something a broad index move can explain: AEP, CMS Energy and FirstEnergy don't sit in any major tech index that could have pulled both portfolios in through a passive fund. For Soros this is under two percent of book on each name, a tactical position, not the flagship-scale bet it is for Thiel. But it's the same list of tickers, and it's not the kind of list that gets picked by accident.

X-Energy is a case of its own: the smallest position in Thiel's book (0.88 percent) turns up at ARK Investment Management nearly thirty times larger in dollar terms, $99.4 million, 0.65 percent of book. A nuclear startup with not a single working reactor is held at once by a former PayPal co-owner, a legendary currency trader, and a growth-stock fund manager, three investors who rarely agree on the same name.

Renaissance Technologies, the largest of the five funds we checked ($72.6 billion in disclosed holdings), holds DTE Energy, FirstEnergy and Vistra, plus sizable positions in bitcoin miners turned AI hosts: Core Scientific ($84.9M), Hut 8 ($34.4M), MARA Holdings ($15.4M), Riot Platforms ($25.4M), and Strategy Inc, the renamed MicroStrategy, at $222.0 million. Renaissance is a quant fund that doesn't give interviews about its theses; the mere presence of this list in its book is its own evidence that the energy shortage and miners' shift into leasing AI capacity have already made it into models trading the market algorithmically, not through a manager's conviction.

Vista Energy is the one Thiel Macro position none of the five checked funds also held. The Argentine shale bet stays Thiel's own call, one no other large fund has repeated.

A timeline of the same thesis

*Source: filing dates from SEC EDGAR (each fund's filing index); tweet dates from tweet metadata via API. The individual miner-to-hyperscaler deals (TeraWulf/Anthropic, IREN/Microsoft, Cipher/AWS) are dated only as "the last two years" in open sources, not to a specific day, so they're left out of this timeline.*

The funds filed independently and at different points within the same week; the regulator doesn't require simultaneous filing. But over that same seven days, four separate voices on X, including the CEO of the company that sells the chip, converged on the same thesis, none of them aware of Thiel Macro's or Soros Fund Management's positions at the time they posted: both filings only became publicly visible on August 14, after the market closed.

An eight-year queue

The reason the outlet became a scarce commodity isn't visible in Thiel's filing. It's visible in the grid interconnection queue.

According to US grid operators, roughly 2,600 gigawatts of new generation and storage applications were sitting in interconnection queues as of early 2026, more than the country's entire installed generating fleet. In PJM, the largest US power market by number of customers, the average time from application to commercial operation has grown from under two years in 2008 to more than eight years in 2025. In Texas, under grid operator ERCOT, the large-load queue has reached 410 gigawatts, and 87 percent of that queue is data centers. At utility Dominion, wait times for a new substation connection now exceed 36 months, and industry sources put the wait for a 100-megawatt connection at roughly seven years.

Eight years isn't an abstract figure from a regulator's report. It's the gap between the day a company decides to build a data center and the day power can physically reach it. Everything that happens in that gap, leasing existing capacity, buying already-operating power plants, and signing nuclear-operator contracts that bypass the shared grid entirely, is the trade Thiel made.

Miners turned landlords

While Thiel's fund was assembling a portfolio of wires and generators, bitcoin miners spent a second straight year running a similar but mirrored trade: selling not electricity, but a site already wired into the grid.

By industry trackers' count, public bitcoin miners have contracted roughly 7 gigawatts of power to hyperscalers and specialized AI clouds over the past two years, across 19 deals worth more than $135 billion. That's under a quarter of the industry's announced 30-gigawatt construction pipeline, meaning the trade is still far from saturated.

The largest contracts by name:

By CoinShares' estimate, the share of public bitcoin miners' revenue coming from AI and high-performance computing could reach 70 percent by the end of 2026, meaning most of them will stop being mining companies in the old sense and become landlords of grid-connected land with ready-built cooling and power infrastructure.

The flip side of the same shift is how miners are funding it. Marathon Digital (MARA), the largest public bitcoin miner, sold about 23,093 bitcoin for $1.63 billion in the first half of 2026 (an average of roughly $70,631 per coin), leaving 35,577 BTC in reserve (about $2.08 billion as of the end of June). On August 4 the company took on another $600 million loan from Coinbase and Two Prime, collateralized by 18,750 bitcoin, partly to fund the Long Ridge acquisition. In other words, part of the capital miners are using to fund their move into AI infrastructure comes from selling and pledging the very asset these sites were originally built to mine.

The gap is clear: Thiel is buying electricity as a raw resource. Miners are selling an already-built grid connection as a finished product. Between those two trades sits all the value the AI industry is currently paying to keep its chips from sitting idle.

The memory bottleneck

Even if electricity were instantly available everywhere, the next constraint would be memory. HBM, high-bandwidth memory packaged next to an AI accelerator in a single die, has become the scarcest component in the whole chain. SK Hynix, the largest maker of that memory, held roughly 57 percent of the global HBM market by revenue as of Q3 2025, and by the company's own account, had sold out its entire 2026 output before the year began. SK Group chairman Chey Tae-won publicly puts the AI memory shortage horizon at 2030.

The cause is production physics, not just demand: to deliver the same amount of usable data, HBM needs roughly 300 percent more silicon wafer capacity than standard DDR5. HBM also carries three to five times the margin of standard memory, so manufacturers are shifting fabs that used to make ordinary chips for PCs and phones, which is why the shortage is spilling over into consumer electronics too. SK Hynix has announced plans to double its production capacity over five years, but by its own admission, no meaningful relief should be expected before 2026-2027.

The packaging bottleneck

Even finished memory and a finished chip don't become an AI accelerator until they're joined in one package, and that's where the bottleneck has shifted to TSMC. CoWoS packaging, the "bridge" between a logic chip and a memory stack that delivers the data-transfer speed AI workloads need, went from roughly 35,000 wafers a month at the end of 2024 to about 75,000 at the end of 2025, targeting 125,000 to 130,000 by the end of 2026: nearly a fourfold increase in under two years, and still not enough. TSMC's CEO said in June 2026 that CoWoS capacity remains "extremely tight and sold out through 2026," and industry analysts say the company's three packaging lines are booked through 2027, with order lead times of 52 to 78 weeks.

That builds a chain of three sequential shortages: electricity, up to an eight-year wait for a connection; memory, sold out a year ahead with a shortage forecast through 2030; chip packaging, booked through 2027. Any one of the three is on its own enough to stall the whole pipeline, and none of them is solved by buying one more Nvidia GPU.

Compute as collateral for debt

A fourth layer sits on top of all this: the money needed to build any of it. Nvidia, alongside the world's largest asset managers, is assembling independent financing platforms to mobilize more than $500 billion for AI infrastructure, not through a funding round in itself, but by turning its own compute capacity into collateral for long-duration debt funded by pension funds and sovereign wealth.

Other market observers confirm the same mechanism from a different angle: the four largest tech companies, Alphabet, Amazon, Meta and Microsoft, have racked up more than $2.4 trillion in future off-balance-sheet obligations, leases that haven't started yet plus contractual commitments to buy capacity, power, servers or cloud services. Years of compute demand are already locked in on paper, and walking away from those obligations is far harder than trimming next quarter's capital budget.

On August 17, 2026, this exact shortage got a name from the man who runs the world's biggest chipmaker. In an X article titled "Securing the Infrastructure of Intelligence," Nvidia CEO Jensen Huang coined the term LPS, land, power and shell, and called it the next strategic resource after chips, packaging, memory and networking. Nvidia is partnering with SB Energy to secure LPS capacity at the PORTS-Pike campus in Portsmouth, Ohio: an initial 4.25 gigawatts over 20 years, with an option to extend to 8 gigawatts. The tenant will be OpenAI. Each generation of Nvidia equipment on the site, per Huang, represents roughly 1.5 million GPUs, or $150 billion to $200 billion in Nvidia revenue. OpenAI's total Nvidia commitments run to roughly 12 gigawatts of compute, with an option to expand to 16 gigawatts, an opportunity worth roughly $600 billion in Nvidia revenue through 2030.

Huang pre-empted the question Barraud asks below. "Is this circular financing?" he asks in the piece, and answers it himself: "No. OpenAI will pay the lease." Nvidia's guarantee covers only part of the lease and power payments, not the full cost of the site, and phases in as data centers come online between 2028 and 2030. The argument is the same one a16z already made about CoreWeave: Nvidia compute is versatile and portable between tenants, so this isn't a circular deal with one customer but an asset that can be resold if OpenAI walks away.

One cloud provider's own revenue confirms this without any financial engineering at all. Nebius (NBIS) reported AI cloud revenue of $582 million in its latest quarter, up 514 percent year over year, with 5 gigawatts of capacity already under contract. According to a broker note on the AI infrastructure market, some of these contracts are structured so the customer prepays up to half of the cloud's own capital spending up front, meaning demand finances the build before it's even finished, not the other way around.

Cuban's line pulled in more than 950 replies within two days, and tellingly, the very first commenters asked whether he meant cryptocurrency wearing a new name. That's a telling question: on this 2026 market, the line between "infrastructure for computing" and "a speculative asset trading on expectations" isn't obvious to everyone, including the people arguing about it.

That demand pressure isn't abstract. Companies like B3, whose B3IQ product is already used by faculty and research teams at Stanford, NYU, Penn, UChicago, Dartmouth, Waterloo and the University of Hawaii, are offering a third path to compute, neither renting from a centralized cloud nor buying hardware outright, but shared ownership of capacity. Demand for that kind of model is rising in lockstep with how cheap it's getting to run a software agent: according to Y Combinator president Garry Tan, a team of two or three people with a few hundred "skill files" can already go from zero to $15 million in annual revenue in four months, and it's that economy, where labor is replaced not by one big chatbot but by an army of parallel agents, that's pushing compute demand faster than the supply of outlets, memory and packaging can grow.

The geography of demand is also spreading past the US. Armenia opened the first phase of the Firebird AI campus in 2026: roughly $500 million invested and 6,144 Nvidia Blackwell GPUs delivering up to 110.6 exaflops, while Kazakhstan is building a "Data Center Valley" in Ekibastuz at 125 megawatts alongside operator Kazakhtelecom, and laying a Trans-Caspian fiber link to Europe. The logic is the same as Thiel's portfolio: hunt not for spare GPUs, but for spare electricity, and build the data center wherever the outlet isn't stuck in a queue.

Mapping the chain: where to look

Market analyst Shay Boloor laid out AI infrastructure as eight layers in mid-August 2026, from the power grid to the end agent, naming a handful of public tickers per layer. We checked that against our own reporting above and kept only the links in the chain this piece has backed with numbers, not general description. This is a map of where the shortage physically sits, not a recommendation; the desk makes no return promises about any name below, and part of this list is already trading on expectations rather than executed contracts.

We don't have a direct answer to "where's the next multiple," and we're not publishing one: the desk's rule is not to promise returns and not to name a venue or ticker as a guaranteed way to get them. All this article's data can say: while all seven nodes in that table are physically constrained at once, the cost of that constraint falls not on whoever owns the chip, but on whoever owns the outlet, the memory, or the packaging line, and those three nodes make up three-quarters of Thiel Macro's portfolio.

What the other side says

The argument here isn't over whether compute demand is growing; nobody quoted above disputes that. It's over who ends up holding the bag if that demand slows down faster than the infrastructure gets built.

Barraud, the macroeconomist quoted above, states the risk directly: the market is now building a financial system that makes the boom cheaper to keep going in the short run and more dangerous when it reverses. With contracts signed years in advance, walking away from them is nearly impossible, so the capital-spending cycle keeps running even if some participants want to slow it down. But if AI revenue doesn't grow as fast as connected capacity does, data-center utilization starts falling, the price of compute drops, and refinancing infrastructure debt gets harder. The same debt mechanism accelerating construction today would run in reverse once demand turns.

The people on the other side of that argument, Huang, a16z, Cuban, and the desk's own view in the quotes above, don't deny that risk outright, but point to what makes this cycle different from previous ones: the contracts are signed against a physically limited resource, electricity, memory, chip packaging, not against the speculative growth of an asset's price, and that supply can't be ramped up quickly in response to falling demand the way general-purpose cloud servers can. In other words: even if some of the models and startups at the top of the chain never pay off, the outlets, wires and packaging lines stay physical assets with alternative uses.

What we can't see

The Thiel Macro portfolio we broke down line by line is a quarterly snapshot from June 30, 2026, filed with the usual 45-day delay built into Form 13F. It doesn't show intra-quarter trades, options, or positions the fund may have closed after that reporting date. Form 13F also doesn't capture Thiel's private investments through Founders Fund and other vehicles, so his real exposure to energy, chips or AI could be wider than what's visible here, or, just as easily, offset by positions we can't see at all.

We have no statement from Thiel or Thiel Macro about the reason for the trade; the entire portfolio section is built on the fact of ownership, not the owner's explanation. The link to the broader energy-scarcity thesis is our own reading of the portfolio's composition, not a restatement of his motives.

The contracted-revenue figures for bitcoin miners, TeraWulf, IREN, Core Scientific, Cipher, are company-reported totals for multi-year contracts, not audited revenue actually received. Some of that money may only materialize years from now, and only if both sides hold up their end of the contract.

The HBM and CoWoS utilization figures are industry observers' estimates and TSMC's and SK Hynix's own public statements, not our own measurement of any production line; we can't independently verify the exact utilization of a specific plant.

The numbers on the Nvidia, SB Energy and OpenAI deal come from Nvidia's own CEO, in a piece written by Nvidia, not an independent check of the contract terms. The $150-200 billion per generation and $600 billion through 2030 estimates are Nvidia's own forecast, not a third-party-confirmed figure. A company that sells chips has an interest in showing demand for them as large as possible.

The five-name overlap between Thiel Macro and Soros Fund Management is a match in portfolio composition on a single reporting date, not proof of coordination between the two funds: 13F doesn't disclose who bought first, or whether the two managers ever spoke at all, and the list of liquid, publicly traded grid-utility names isn't so long that a two-fund overlap is statistically implausible. Each of these positions is under 1.3 percent of book for Soros, a secondary bet in his portfolio, and we didn't check whether the fund held the same names a year earlier as part of an ordinary utility-sector rotation.

What would have to be true for us to be wrong

Here is what would change our mind about the "the bottleneck is the outlet, not the chip" thesis:

• Grid interconnection queues in key US markets, PJM, ERCOT, Dominion's territory, would have to shrink rather than grow over the next 12 months.
• TSMC's CoWoS utilization and SK Hynix's and Samsung's HBM output would have to come back into balance with demand: order lead times would have to fall well below the current 52 to 78 weeks, rather than staying there through 2027 as the industry itself claims.
• At least one of the major mining-to-AI contracts, TeraWulf/Anthropic, IREN/Microsoft, Cipher/AWS, would have to be canceled or materially cut back without a comparable replacement deal.
• Hyperscaler data-center utilization (Amazon, Microsoft, Google) would have to start dropping noticeably even as installed capacity holds steady or grows, a direct sign that compute supply has outrun real demand.
• In the Q3 2026 13F, due to be filed in mid-November, both Thiel Macro and Soros Fund Management would have to exit most of their power-utility positions at the same time and without replacing them with comparable energy names; that would mean the overlap in "Who else holds the same names" was a one-off, not durable positioning by large capital.

If all of that happens at once, the "the bottleneck is the outlet, not the chip" thesis stops holding, and Thiel Macro's portfolio should be read as badly mistimed rather than structurally right.

The TT desk's read

We're now adding power-utility and chip-manufacturing names, Vistra, the US grid utilities, X-Energy, SK Hynix, TSMC, to our own TradFi coverage alongside AI clouds we already track like Nebius, not because it's guaranteed money, but because four independent trails, Thiel Macro's 13F, the same list of names at Soros Fund Management, miners' contracts, and memory and chip-packaging producers' own reporting, currently point to the same physical limit, and it's easier to watch it on one panel than to reassemble it from scratch every time a new quarterly filing lands. We're also setting up a quarterly 13F check across five or six large funds against this specific ticker list; next time, "Who else holds the same names" won't be a one-off breakdown but a running snapshot.

The number we'll keep watching: TSMC's CoWoS order lead time (currently 52 to 78 weeks) and the share of public bitcoin miners' revenue coming from AI and high-performance computing (currently on track for 70 percent by end of 2026, per CoinShares). If the first number starts falling while the second keeps climbing without a matching rise in contracted capacity, that's the first sign supply is starting to outrun demand rather than the other way around.

Check one specific number behind our back: the 52.8 percent share of Thiel Macro's portfolio in power utilities, calculated from five rows, AEP, DTE, FirstEnergy, CMS Energy and Vistra, divided by the $418.7 million total. You can recompute it directly from Thiel Macro's own SEC EDGAR filing, with zero intermediaries.

Keep reading

Crypto Capital Flows Analysis · All notes