The central bank of Jensen, Claude goes into the lab, and the falling price of benchmarks

Nvidia CEO Jensen Huang spent this week on the front foot. In a long interview with Ezra Klein for the New York Times, Nvidia's chief executive described AI safety as an engineering problem. If a lab cannot contain its experiments, he said, "we have to shut the labs down". He said Geoffrey Hinton's 10% estimate of societal collapse was "not grounded on science". He also argued that scaremongering is doing real damage to the industry, and spent time on how data centres can win over the communities around them.

Klein opened by noting that Nvidia is worth $5.4 trillion. He also said it has produced 15 cents of every dollar the US stock market has returned since 2023. A company of that size has a financial stake in how the public feels about AI. Over the past two years Nvidia has also become the industry's lender of last resort. Its loans depend on something Nvidia's engineers can't fix: how quickly grids, planners and local communities allow data centres to be switched on.

Huang told Klein that Nvidia had invested about $100 billion across the industry. The Economist, in a briefing called "Nvidia is the central bank of AI", counted more than $70 billion pledged to startups, close to his figure. His number leaves out the rest: about $300 billion of potential liabilities to customers. These include promises to rent back unused capacity from neoclouds such as CoreWeave, support for up to a quarter of the residual value of chips leased through Wall Street partners, and a guarantee of the land, power and shell at SB Energy's Ohio campus, leased to OpenAI for 20 years. SemiAnalysis estimates that the neocloud programme alone creates about $59 billion of obligations per gigawatt, and that most of the capacity Nvidia backstops is not yet built.

But all is not well with the projects meant to turn those promises into working data centres. Oracle's first-quarter results showed a contract backlog of $664 billion, more than half of it from OpenAI by analysts' estimates. But free cash flow was negative $5 billion, interest costs rose 55% and the company sold $20 billion of new shares. S&P cut it to one notch above junk in July. This week Oracle sent a force majeure notice to Blue Owl's STACK Infrastructure over Project Jupiter, the New Mexico campus being built for OpenAI. Oracle plans to run it entirely on up to 2.45GW of Bloom Energy fuel cells on site, one of the largest fuel-cell installations in the world. The notice cites potential delays in securing that power, and lets Oracle delay payments if the site misses its 2028 date. About $18 billion of loans tied to the site were already quoted at around 90 cents on the dollar. Oracle says Jupiter remains on schedule.

Project Jupiter under construction in June. Epoch AI projects the equivalent of about five million H100 chips there by late 2028, costing about $66 billion. Image: Epoch AI.

In Essex, Nscale's Loughton site, announced last year as Britain's largest AI supercomputer, was due to open for Microsoft in 2027. UK Power Networks has reportedly told Nscale the power may not arrive until the early to mid-2030s, because the transmission network cannot yet carry it. As we covered in The $50 million megawatt, the UK grid is heavily oversubscribed. Ofgem now counts 315 data centres queuing for 73GW of grid connections, against national peak demand of 45GW. Days after filing for a New York listing, Nscale raised $3.36 billion, $1 billion of it from Nvidia. Its prospectus discloses that management once doubted it could continue as a going concern. Nvidia sells Nscale its chips, owns part of it and now helps fund it.

The UK's grid queue sits beside a far larger US build-out. Source: Cushman & Wakefield, analysis by Computer Weekly.

Central banks can print money in a crisis. Nvidia's nearest equivalent is $56.6 billion of cash and marketable securities and tens of billions of dollars of cash generated every quarter. But its guarantees are written against capacity that is built, powered and ready to deliver. CoreWeave's filing makes Nvidia's promise to buy unsold capacity subject to "delivery and availability of service requirements". In Ohio, Nvidia guarantees payment for power, not the arrival of power. If a site can't be energised, the backstop has nothing to pay out on. The loss then falls on the developer, the lender or whichever tenant signed the weakest contract.

An engineer who installs electrical systems in data centres told us the pressure to hit schedules is intense, because chips that aren't running are losing value. The accounts don't show it at first, because depreciation only starts when equipment goes into service. The economic loss begins earlier. Interest starts the day a loan is drawn, but Nvidia releases a new architecture roughly every year; Hopper, Blackwell, Rubin, and so on. The buildings age as well. Rubin racks draw 190 to 230kW. Rubin Ultra, in Nvidia's Kyber racks, is expected to draw 600kW and run on 800V DC power from 2027. A hall designed in 2025 around conventional AC power and then delayed by a year could open with an electrical design a generation behind the chips it is meant to house.

Projects wait in the same connection queues and for the same transformers and switchgear, so their delays tend to arrive together. They face the same local opposition, which blocked or delayed nearly $200 billion of US projects in the first half of this year. Chip-backed debt is increasingly split into tranches in the style of collateralised loan obligations, a structure that works best when defaults are independent. Projects waiting on the same grid bottlenecks are not independent. The Bank of England judged the stock of AI-related debt modest in July. Counting how many financed projects wait on the same substations might tell regulators more.

Older chips are holding their value, as Huang says. SemiAnalysis's price index puts an H100 at about $2.80 an hour, roughly where it stood a year earlier. CoreWeave has signed an A100 contract that runs to 2029, nine years after the chip launched. One-year contract prices for H100s rose about 40% between October and March. The reason is a shortage of powered capacity. The grid constraints delaying Jupiter and Loughton are the same ones keeping existing chips busy and expensive. The shortage that holds up the value of Nvidia's collateral is the same one putting its guaranteed projects at risk.

In the interview, Huang put the cost of a gigawatt AI factory at about $50 billion and its rental income at $40 to $50 billion a year. On his own figures, a one-year delay to a gigawatt site means tens of billions of dollars of lost revenue while interest keeps building up. His campaign against the doomers also protects Nvidia's loan book. Its exposure grows with every megawatt it backs, and those megawatts depend on planning permission, grid upgrades and local consent. Public confidence in AI is one of the few things behind his guarantees that he can try to influence, and he is using the largest platforms available to do so. Central banks use speeches in a similar way, managing expectations to protect the value of what they lend against. The difference is that Nvidia also sells the product its loans are secured on.

Takeaways: Huang has a case that many AI safety problems can be treated as engineering problems, and older chips are holding their value as he says. But Nvidia is now also a lender, with about $300 billion of potential liabilities tied to megawatts it cannot switch on itself. Its own balance sheet can absorb a lot. The neoclouds, developers, lenders and Oracle beneath it are less protected, and power delays tend to hit them together. UK organisations reserving compute with Nvidia-backed providers should ask where the power comes from, when it is contracted to arrive and who pays if it doesn't.

Anthropic now runs its own biolab, a demonstration of how broad its ambitions have become, and on Wednesday it published its first discovery. About 950 Claude agents spent 21 hours going through more than 200,000 reverse transcriptases in public genome data. They flagged 3,500 candidate systems and wrote reports on the 20 strongest. One agent noticed a line of evenly spaced DNA repeats beside an enzyme in a virus that infects bacteria. That arrangement looks like CRISPR. Anthropic has called the system ART. Its human scientists have shown that the repeat array is converted into short RNAs, but nobody yet knows what ART does.

Anthropic's scientists decide which of Claude's ideas go to the bench.

The approach is interesting in itself. Anthropic has not given Claude control of lab robots. Claude reads data at large scale and proposes hypotheses, and people at the bench test the few that make it through. The team also records which of Claude's ideas its scientists decide to pursue. Those decisions are then used to shape how Claude ranks future ideas. The company is trying to reproduce scientific judgement, which is the skill every lab has least of.

“Eventually it may even be possible for Claude itself to safely perform the experiments by autonomously controlling lab equipment, with appropriate safeguards in place, but we aren’t doing that today.”

Anthropic reported a weakness in the method itself. The team ran ten more campaigns with the same setup, and none found the ART array again. None of the later agents looked at the right stretch of DNA. The discovery depended on a choice made by a single agent. For now, running more agents mainly increases the chance that one of them looks in the right place.

Many readers will be uneasy that a company warning about AI-enabled bioweapons now runs its own biology lab. That concern is reasonable. Anthropic turned on its strictest safeguards for Claude Opus 4 because it could not rule out that the model might help someone with basic scientific training build a biological weapon. The new lab works only at biosafety levels 1 and 2 and handles no human pathogens. Dario Amodei has said Claude may one day run experiments itself "with appropriate safeguards". Anthropic has not yet described those safeguards in public. The bigger risk comes from the skills the lab is building. Searching genomes for unusual systems and designing proteins that bind to specific targets can help medicine, and the same skills can be misused.

Anthropic is one of several groups building AI-driven labs, and each has made a different bet on where science is slowest:

• Edison Scientific runs no lab at all. Spun out of the non-profit FutureHouse in 2025 with $70 million, it sells Kosmos, an AI scientist that reads around 1,500 papers and writes tens of thousands of lines of analysis code in a single run. Its bet is that most of the answers are already in published work and in the data that drug companies hold.
• Lila Sciences, launched by Flagship Pioneering, the firm that created Moderna, is building fully automated labs across biology, chemistry and materials, where AI designs and runs the experiments and learns from the results. It has raised about $550 million and describes its goal as scientific superintelligence.
• Periodic Labs was founded last year by Liam Fedus, one of the creators of ChatGPT, and Ekin Doğuş Çubuk, who led materials discovery at Google DeepMind, with a $300 million first round. Its founders argue that the data needed to teach a model physics and chemistry does not exist online and can only come from running experiments. One early target is superconductors that work at higher temperatures.
• Isomorphic Labs, spun out of Google DeepMind and led by Demis Hassabis, works only on medicines. It raised $2.1 billion in May, partners with Novartis, Eli Lilly and Johnson & Johnson, and expects its first AI-designed drugs to reach clinical trials by the end of this year, a year later than first planned.

Anthropic sits between Edison and the automated labs. AI does the searching at huge scale, and people do the lab work.

Across all of these approaches, finding candidates is getting cheaper and understanding them is not. ART took less than a day to find, and working out its function will take months of lab work. Organisations that can pay for hundreds of millions of tokens and experienced scientists to review the results will move ahead. At Harvard, every student and member of faculty can request up to $2,500 a month of Claude, even as the university's research funding is under threat and it lays off staff.

Takeaways: Anthropic's lab has shown that AI agents can find new biology in data that people have already studied closely. So far this happens unreliably, and humans still do every experiment. The open questions are who checks the findings, how quickly the checks can be done, and what protects the process from misuse. Answers to those questions will decide who benefits from AI-driven biology, more than the number of discoveries does.

Our chart this week comes from Epoch AI. It estimates that the cost of reaching a fixed level of AI performance has fallen by about 47% per quarter since 2023, or roughly 13x per year. That is faster than the declines measured for electricity, lithium batteries, DNA sequencing and computing. Scoring 75% on the GPQA Diamond science exam cost around $0.30 per question with o3 in January 2025. Under 18 months later, GPT-5.6 Luna matched it for $0.0004.

The rate varies by task. Maths benchmarks show the steepest falls and games of skill the slowest. On SWE-bench Verified, which tests the software engineering work businesses increasingly give to agents, the cost has fallen by about 27.5% per quarter since mid-2024, or around 3.6x per year.

Benchmarked intelligence is clearly getting much cheaper. Complex outcomes often need far more tokens, because agents working on long tasks reason, retry and call tools many times over. The cost of a finished piece of work can therefore fall more slowly than the cost of answering a single benchmark question.