The Same Equation, at Every Scale

On July 20, a Beijing startup called Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model that matched GPT-5.6 and Claude Fable 5 on coding and reasoning benchmarks ^[1][2]. Within 48 hours, the company acknowledged the strain.

To protect the experience of existing subscribers, we're temporarily pausing new subscriptions and prioritising compute for current members. — Moonshot AI

It paused new subscriptions ^[2]. The episode read like a startup scaling problem — the kind a fresh funding round is supposed to solve. But Moonshot had already raised $3.9 billion in six months at a valuation above $20 billion, with Alibaba and Tencent among its backers ^[3]. After K3 launched, annual recurring revenue surged from $200 million to $300 million in two months ^[1]. The company is now seeking up to $2 billion more and racing toward a Hong Kong IPO at a $30 billion valuation ^[1][2]. Even revenue growth that would make any software company a public-market success cannot keep pace with the cost of serving a frontier model at scale. The market appeared to draw the same conclusion, fast. When K3 landed, the Nasdaq dropped 2.6% and South Korea's KOSPI — already in bear-market territory from an earlier AI sell-off — fell further, down nearly 30% from its June peak with multiple trading halts ^[2]. This was not a reaction to collapsing chip demand. Samsung had just reported record quarterly operating profits of nearly 90 trillion won, and investors were concluding "that artificial intelligence infrastructure spending by U.S. firms cannot sustain current valuations" ^[4]. The sell-off coincided with a recognition that model parity from a Beijing startup meant competitive moats were thinner than assumed — and that the firms buying the chips might not be solvent enough to keep buying them. Multiple forces were at work; no single event can be isolated as the sole trigger. Now widen the aperture. The same equation runs through Anthropic, an order of magnitude larger. The company closed a $30 billion round at a $900 billion valuation in May, with $45 billion in projected 2026 revenue ^[5]. It still expects to lose money for the full year. It has locked in a $1.25 billion monthly compute deal with SpaceX through 2029 ^[5]. CEO Dario Amodei put the strain in plain terms.

I hope that 80-times growth doesn’t continue because that’s just crazy and it’s too hard to handle. — Dario Amodei

Months earlier, Anthropic had already been throttling Claude users during peak hours, reducing session limits for roughly 7% of users without formal notice ^[6]. The bind is not a startup's growing pain. It is structural. Then OpenAI. The largest frontier lab by revenue is seeking $100 billion in new capital and does not expect to reach profitability until 2030 ^[7]. Amazon is negotiating a $50 billion chips-for-equity arrangement that would grant OpenAI access to proprietary Trainium and Inferentia chips via AWS. One source close to the talks described the arrangement.

Together with OpenAI we’re driving the AI revolution to the next stage. — Masayoshi Son

The market leader is bartering ownership stakes for compute access because it cannot fund infrastructure from model revenue alone. The counterfactual is visible in the earnings reports. Amazon and Alphabet are projecting combined 2026 capital expenditure of roughly $380 to $390 billion, funded directly from cloud operating cash flow: AWS revenue rose 28% year over year to $37.6 billion, and Google Cloud grew 63% to $20 billion with a $460 billion backlog ^[8]. Alphabet's CFO was explicit about the trajectory.

revenue from products built on our gen AI models grew nearly 800% year over year. — Sundar Pichai

These are the entities that can self-fund frontier infrastructure — not from model revenue, but from the cash thrown off by the cloud businesses that sell compute to everyone else. The absorption deals are already taking shape. Amazon's $50 billion chips-for-equity negotiation with OpenAI is one form. Another is Tencent and Alibaba — Chinese hyperscalers with combined AI investments exceeding $5 billion annually — entering advanced discussions to invest in DeepSeek and integrate its models into their cloud platforms as a "national champion" ^[9]. DeepSeek, long self-funded through founder Liang Wenfeng's quant hedge fund, broke that policy to raise $3 to $4 billion in its first external VC round at a $45 to $50 billion valuation, with China's state-backed national AI fund negotiating to lead ^[3]. Even its attempt to develop custom inference chips to reduce dependence on Nvidia and Huawei requires raising $7 billion at a $52 to $59 billion valuation ^[10]. The cost-reduction strategy itself demands more capital, not less. These absorption deals are early-stage and directional, not a completed outcome. The sell-off that coincided with K3's launch is a correlation, not a causal chain proven link by link. But the pattern is consistent across three orders of magnitude: a $20 billion Beijing startup, a $900 billion San Francisco lab, and the market leader seeking a $1 trillion valuation all face the same arithmetic. Whether that arithmetic hardens into an endgame depends on something no lab has yet demonstrated: the ability to grow model revenue faster than the infrastructure cost of serving it.