The Investor

What changed: Strategic buyers reset the exit ceiling on distribution assets. Labs pushed compute risk six years forward onto balance sheets you do not underwrite, and China's premium multiple now sits behind three impaired exit routes.

Why it matters: A record print prices what the buyer feared. Consideration mix and the half-life of the purchased layer decide whether adjacent marks can follow it.

What to do: Consideration structure in writing does most of the work here, since an asset swap is not a cash comp. Anchor-tenant contract terms carry every neocloud position, and China exposure re-underwrites on secondaries-only liquidity.

The mechanism behind the 1,000x

BoxGroup's outcome is a reserve-strategy story before it is a founder story. A $750,000 first check into Michael Truell in 2022, two follow-ons after that, and Fortune's Term Sheet now puts the fund-level result at roughly $1 billion. A then-principal sourced and championed the check, not a partner. The original pitch was AI for CAD, so the category's largest exit came from a founder who publicly abandoned the thesis he was funded on. Reserve architecture did the work idea diligence usually gets credit for.

Two cautions travel with the print. The $1B rests on a single source familiar with the deal, framed as what should shake out. And SpaceX is private: a price paid partly in the equity of an unlisted, founder-controlled aerospace contractor is an asset swap into another illiquid mark. A government contractor buying developer tooling at this size also invites antitrust and national-security-adjacent review, which lengthens the path to close.

Three prints, one pattern

Nvidia's number is the most legible. January's reported offer was $7B, roughly 47x. Seven months later it agreed at about $12.9B, roughly 80x, and the only disclosed operating change in between was Hugging Face doubling its customer base. That 84% is a chokepoint premium rather than a revenue multiple, paid by the silicon incumbent for the layer sitting above its own hardware, in the same month its largest customer shipped competing inference silicon.

Where the reads diverge, and how to resolve it

One read says exit assumptions reprice and entry math holds; the other says seed basis got worse overnight, as every fund hunts this exact profile and capital rotates toward brand-name pre-seed managers. Both hold at different points in the stack. Strategics just raised terminal value on any asset owning a registry, package manager, hub, or default-install position, even below $200M of ARR, and founders will anchor to 80x in negotiations within days. This is probably wrong, but the discipline that survives contact is a written 15–20x financial-buyer floor with the strategic scenario probability-weighted on top.

The layer is commoditizing while it is being bought

Three companies launched the same aggregation wedge inside 24 hours: AgentSky across agent harnesses, Monid across 1,700-plus tools, Bezalel behind a single MCP URL. Ramp shipped its own router publicly while Stripe was buying one. The savings math is worse than that, because prompt caches are model-specific and switching models mid-session forces re-payment of the full input-token cost. In long-context agentic workloads, which is precisely where routing lives, list-price arbitrage can invert into negative savings.

A record price for a distribution layer prices the incumbent's fear of the layer above its own silicon: $7B in January, about $12.9B seven months later.

Read the ramp, not the average

Forty-five billion dollars over six years averages roughly $7.5B a year of fixed compute cost for one lab, per Bloomberg's disclosure, which is the number everyone will quote and also the least informative one, since real schedules are almost certainly back-weighted and out-year exposure therefore runs above the average. Kioxia's program averages about $5.2B a year on economics that lean on a Japanese subsidy nobody has confirmed. Log it as a named monitoring item: a denial slips capacity, and every hardware bill of materials in the book gets worse. Kioxia rose 5.0% to ¥52,500 on a capex announcement that in earlier cycles would have been sold as a coming glut. Memory now prices like a bottleneck.

The Nscale template is dangerous

Nscale went from startup to systemically relevant counterparty on the strength of one customer, which is an interesting deal puzzle before it is anything else. Viability rests on three things outside its control: the lab's solvency, the lab's continued demand, and its own ability to deliver a very large facility on schedule. Call it concentrated credit risk plus construction risk dressed as contracted revenue. The treatment is a concentration discount, and the diligence asks for minimum volume commitments, termination rights, credit support and ramp schedules before anyone applies a multiple to any of it.

The credit channel is the genuinely new part

A revolver above $10B, tiered at roughly $1.25B for the bulge bracket, about $1B for tier two and $750MM or less for everyone else, is underwritten on relationship logic: banks are buying franchise access to an AI listing with balance sheet. The benign read is that they are paid to be there and the tickets are small against their books, and that read is defensible. It also puts the offering nearer than secondary pricing implies, because pre-listing revolvers do not sit idle, and it puts correlated AI exposure inside bank capital, so a model-layer revenue miss travels through bank balance sheets into the financing cost of every company held downstream.

Delete compute deflation from the 2027 model

Amazon tripled cumulative Nvidia commitments to roughly 3 million GPUs in five months while its own accelerators run at a $25B annualized rate against $225B of lab commitments, about 9x forward coverage, and it still paid a tens-of-billions figure to a company it competes with. At this level of demand, vertical integration is additive. Blackwell Ultra, Rubin and Rubin Ultra do not land in AWS until the 2027–2028 window, so efficiency between now and then comes out of architecture.

Where the obvious cost-relief trade gets complicated

Fixed compute obligations at the lab layer create durable demand for inference optimization, routing and small-model substitution, which is the trade everyone will find first. The complication is that the efficiency is open source: paged key-value cache management delivers two to four times more users per GPU at roughly 96% cache utilization, and anyone can have it. Prompt-cache pricing has standardized at 1.25x base rate to write and 0.1x to read at both frontier providers, with break-even near 1.3 requests inside the cache window, which turns cached-read share into a management-quality signal rather than a vendor-negotiation outcome, and leaves retrieval-heavy architectures locked out of the largest lever in their own cost line. This thesis could be wrong if pricing moves, but until it does, underwrite cost relief on workflow lock-in and contracted offtake.

One tell worth logging: the most sophisticated commentary in this capex cycle has moved from benchmarks to depreciation schedules and debt covenants. That argument is circulating unquantified, which is exactly why doing the arithmetic in-house is worth something.

Anchor-tenant backlog is concentrated credit risk and construction risk wearing the clothes of contracted revenue.

The pricing mark, and the revenue that isn't behind it

Two numbers do the arguing in the Dimension Capital letter obtained by The Information, and they point in opposite directions. Leading Chinese labs trade at revenue multiples five to ten times higher than US frontier labs. DeepSeek, the most-adopted Chinese open-weight brand in the West, generated roughly $70M in the first seven months of 2026, against tens of billions annualized at Anthropic and OpenAI. So a premium multiple sits on a revenue base two to three orders of magnitude smaller. When the firm actively marketing China field expertise writes that the climate is meaningfully frothier there than in the West, the bull case has helpfully labeled where the risk lives.

Exit mechanics are the actual underwriting problem

The multiple debate is the less interesting half. All three liquidity routes are impaired at once, and that is what should change the diligence template.

The sharpest detail sits inside the Manus outcome. The company had already relocated to Singapore, and that shielded it from neither a US Treasury review of its investor's stake nor Beijing's veto. Redomiciliation is theater. Enforcement follows founders, intellectual property, and operations. Its investor did distribute returns before the deal died, which is timing luck wearing the clothes of a repeatable playbook.

Where the adoption actually monetizes

Chinese open-weight models cleared a Western adoption bar that an entire prior generation of Chinese enterprise software never got near, and then failed to capture the revenue that adoption generates. American developers run the models. American serving companies bill for them, because they can supply what Chinese vendors structurally cannot: enterprise trust, compliance posture, data residency, service-level commitments. Modal, Baseten, and Fireworks are the commercial counterparty of record, and they are still priced as commodity compute.

Two independent data points from this cycle corroborate the commoditization half of the letter and leave the pricing half entirely alone. Z.ai revealed its stealth model as GLM-5.3-Flash and released the weights the same day, with day-one support across the major open serving stacks. And Visa's published vulnerability harness treats an open-weight Chinese model as interchangeable with the frontier labs at a flagship account. Capability is depreciating as a differentiator on a timeline measured in hours.

The operating consequence

US firms have converged on reconnaissance without capital, which looks like a week in-country here, ten days there, mutual funds making the trip, nobody deploying. The letter's own thesis is explicitly conditional on regulatory risk being assuaged, which is option-buying language: build relationships now, deploy only on a policy catalyst. There are versions where this resolves differently. A policy thaw arrives and the relationship-builders look prescient rather than idle. Or a Chinese lab finally converts Western developer share into Western enterprise revenue and the multiple stops looking absurd. Or the serving layer consolidates and the margin ends up with whoever owns distribution instead. This is probably wrong in one of those directions, but the view is that the transferable value is the intelligence, not the cap-table entry. How fast Chinese founders route around compute restrictions is the leading indicator for how fast US model advantages decay, and that signal is worth more to a Western book than any China position that could realistically get approved.

China's labs have won Western developers and lost Western revenue, so the asset to diligence is the serving layer standing in front of them.