← Cross-cutsCross-cut · AI labs

Two ways to bet on the frontier

Eight Teardown studies cover the labs building frontier AI — OpenAI, Anthropic, xAI, Mistral, and China’s DeepSeek, Zhipu, MiniMax and Kimi. Read together, they don’t describe one race. They describe two opposing wagers on how to win it: spend the most, or spend the least. Which bet is right turns on a single question every one of these studies independently reaches — whether the model itself is still a moat.

The valuation gulf is the whole story

Line the eight up by price and a chasm opens in the middle. The two American leaders are valued at $852B (OpenAI) and $965B (Anthropic) — roughly 40× and 20× their own unaudited run-rates, both still deeply loss-making. The four Chinese labs sit one to two orders of magnitude lower: DeepSeek’s reported first external round implied ~$10B, Kimi ~$20B, MiniMax ~$4B. Mistral, Europe’s champion, is ~€11.7B. The gap is not mostly about model quality — DeepSeek’s R1 matched OpenAI’s o1 on reasoning, and GLM-5, M2 and K2 top the open-weight rankings. It is about two different theories of what you are buying.

Capital-maximalist vs efficiency-maximalist

The Western bet is capital-maximalist: raise unprecedented sums, buy unprecedented compute, and monetize a lead before it erodes. OpenAI has signed roughly $1.4T of multi-year compute commitments — about 9× its revenue — and its backers (Microsoft, Nvidia, Oracle) are also its suppliers. The Chinese bet is the mirror image: efficiency-maximalist. DeepSeek is self-funded by a quant fund, takes no outside VC, and matched the frontier under a hard export-control compute ceiling by out-engineering it (MLA, MoE, FP8). The whole cluster is a controlled experiment in whether the frontier is bought or engineered — and the Chinese labs have already proven you can get close for far less.

The question they all answer the same way

Here is the convergence that makes this a real cross-cut: every study, Western and Chinese, lands on the same uncomfortable verdict — the model layer is commoditizing. OpenAI’s own teardown concludes its moat sits at the distribution and app layer (~800M weekly users, 4M developers), not the model. Anthropic’s is enterprise and coding workflow lock-in. The Chinese labs make the point louder by giving their weights away for free — DeepSeek’s own line is that “closed-source moats are fleeting.” If the model is not the moat, then the $850–965B Western marks are bets on distribution and switching costs, and the capital-light open-weight labs are bets that neither will hold. They cannot both be right.

Open weights are China’s strategy, not a side-effect

The open-vs-closed split maps almost perfectly onto the geography. The three highest-valued Western labs are closed-weight; every Chinese lab here, plus Mistral, ships open weights under permissive licenses. That is deliberate. For a Chinese lab under a compute ceiling, open weights buy developer gravity, ecosystem reach and a credibility the capital can’t — GLM-4.5 hit 10M downloads in 30 days; K2 outscored GPT-5 on one hard benchmark. Open-weighting is how the efficiency bet distributes: you can’t outspend OpenAI, so you commoditize the thing it spent the most on. The risk, which the studies flag in unison, is that open weights also mean you may never capture the value you create.

Where they split — and the two wild cards

The labs agree the frontier is deflating fast (capability gets ~4× cheaper a year). They split on who survives it. The Western thesis: scale and distribution become the moat once the model isn’t one. The Chinese thesis: in a deflationary market, the low-cost producer wins, and the high-cost incumbents are over-built. Two names refuse the binary. xAI is its own bet — vertical integration (X’s data, SpaceX’s capital, Colossus’s compute) folded into SpaceX, ~$0.5B of revenue against a ~$6.4B operating loss, the most capital-intensive seat of all. Mistral is the European hedge — capital-light like the Chinese labs but Western-aligned, selling sovereignty and on-prem control rather than raw frontier scale. The cross-section’s sharpest read: the demand for intelligence is not in question. Whether anyone earns a durable profit selling it — at any spending level — is.

The cluster at a glance

CompanyModel & standingFunding & backerRevenue / economicsThe bet
OpenAIPrivateChatGPT consumer leader (~800M WAU)$852B post · MSFT/Nvidia/Oracle~$20B run-rate · ~$9B lossCapital-maximalist, closed
AnthropicPrivate~40% enterprise API · coding #1$965B post · Amazon/Google~$47B run-rate · −94% GM (2024)Capital-maximalist, closed
xAIPrivateDistant #3 · ~13.5% US app share~$45B · Musk (X/SpaceX)~$0.5B rev · ~$6.4B op lossVertical integration, closed
Mistral AIPrivateEuropean champion · open + closed~€2.8B · ASML-led €1.7B~$400M ARR (est.)Capital-light, sovereign, open
DeepSeekPrivateOpen-weight · R1 matched o1Self-funded (High-Flyer)API-only · ~$205M ceilingEfficiency-maximalist, open
Zhipu AIHKEX (listed Jan 2026; first foundation-model AI to IPO)GLM-5 #1 open · first AI IPO>¥8.3B · state + BAT; HKEX¥724M rev · ¥4.72B lossEfficiency, open, state-backed
MiniMaxHKEX:00100M2 #1 open-weight · Hailuo apps~$4B · Alibaba/Tencent; HKEX~$79M rev · 73% overseasEfficiency + consumer funnel
Moonshot AIPrivateK2 beat GPT-5 on HLE · open~$20B post · Meituan-led~$200M ARR (est.)Efficiency, open, API-led

Most figures are reported estimates or company-stated run-rates — every lab here is private except the HKEX-listed Chinese names, and none discloses audited frontier economics. As of each study’s stated date (2026-06); see each teardown for sourcing and the full weighing.

The eight studies — and the question each turns on

OpenAIPrivateCan unmatched scale and brand be turned into a durable, profitable business as the model commoditizes, prices deflate, and far larger rivals stack against it?Read the full weighing →AnthropicPrivateCan a safety-branded lab that is ~40% of enterprise API spend justify a ~$965B price on still loss-making, run-rate revenue while dependent on its hyperscaler rivals?Read the full weighing →xAI (Grok)PrivateCan Musk's ecosystem of data, compute and capital be turned into a durable, profitable AI business — or is vertical integration masking weak standalone economics at a froth valuation?Read the full weighing →Mistral AIPrivateCan a capital-light European lab funded ~10x below US rivals stay at the AI frontier — and is its 'sovereign European' identity a durable moat or branding?Read the full weighing →DeepSeek (深度求索)PrivateCan an open-weight, self-funded Chinese lab stay at the frontier under a hard compute ceiling — or does it settle into a cheap fast-follower?Read the full weighing →Zhipu AI (智谱AI / Z.ai)HKEX (listed Jan 2026; first foundation-model AI to IPO)Can a Tsinghua-bred open-weight frontier lab — earning ~85% from low-margin government/enterprise deployment on US-restricted compute, at a triple-digit sales multiple — convert technical standing into a durable, profitable business as DeepSeek and big tech bear down?Read the full weighing →MiniMax (稀宇科技)HKEX:00100Can MiniMax's efficiency-plus-open-weights edge and overseas consumer franchise compound into durable profit before commoditization, the price war, litigation and geopolitics erode them — given it is priced as if they already have?Read the full weighing →Moonshot AI (Kimi / 月之暗面)PrivateCan a technically capable but sub-scale lab convert frontier open-weight models into a durable, profitable business in a price-deflationary market dominated by far larger rivals?Read the full weighing →

This is the kind of reading the Desk does for you

A cross-cut takes eight teardowns and one frontier and asks what they say together. The Desk does the same for the companies you actually own — your thesis, the rivals that move it, and the tripwires that would change your mind.

See the Desk →