$1.22bn went to cutting inference costs in six days

Four rounds in today's investments digest are selling the same thing, and it is not a model. Etched, Olix, Callosum and Velaura AI raised $1.22 billion between them over six days, and each one is a wager that the expensive part of artificial intelligence is no longer training it but running it.

The four rounds

• Etched raised $700 million led by Jane Street, just under a month after a $300 million fundraise, doubling the inference chip startup's valuation to $21 billion.
• Olix, a two-year-old London designer of photonic inference chips, raised $312 million at a $3.3 billion valuation, triple its mark six months earlier, from Fundomo, Arm, Hudson River Trading, Reed Hastings and the UK government's Sovereign AI fund.
• Callosum raised a $100 million seed led by Atomico, with Plural, DCVC and a significant cheque from the UK's £500 million Sovereign AI fund, for software that matches an AI task to the cheapest suitable model and chip.
• Velaura AI raised a $110 million Series A led by Seligman Ventures, valuing the chip designer above $1 billion on technology that cuts power draw in AI data centres.

Two of those marks moved on a clock measured in weeks. Etched's doubled in under a month, and Olix's tripled in six.

The buyers are voting the same way

Google secured the right to acquire up to $12.2 billion of Marvell stock as part of a custom AI silicon partnership disclosed in regulatory filings, sending Marvell shares up 6%. A supplier contract became an equity alliance, which points where the four rounds point: the largest buyers of compute would rather own a second source than keep renting the first.

The scarcity shows from the other side too. Nvidia's most advanced chips are barred from export to China, yet Chinese firms including ByteDance, Alibaba and Tencent have reached GB300 compute remotely through data centres in Thailand, Malaysia and Japan, a loophole Washington is now trying to close. Nobody routes a workload through three foreign jurisdictions for something cheap and plentiful at home.

A state fund on the cost side

The UK's Sovereign AI fund is in two of the four rounds: Callosum's seed and Olix's $312 million. Neither company is building a frontier model. One is designing the chip the model runs on and the other is deciding which chip that should be, which is a public fund taking a position on the operating bill rather than on the frontier.

For Atomico the position is not new. It backed Lace Lithography in March, framing that investment as rewriting the rules of chipmaking. Its early-stage vehicle, Atomico Venture VI, closed at $485 million for predominantly Series A cheques, which is the measure of how large a $100 million seed is for the firm leading it.

The cheapest path may not be silicon

Z.ai released GLM 5.3, an open-weight model it says automates coding and cybersecurity work almost as well as the best publicly available closed models, with full access promised in two weeks after a limited release to trusted partners. A capable open-weight model puts a ceiling on what anyone can charge to serve one, and it needs no fab, no photonic wafer and no routing layer to do it.

That is the risk sitting under $1.22 billion of new capital. None of these four rounds is a claim about what models can do. All four are claims about what it costs to run them, and the ceiling on that number may be set by something given away rather than by something built.