Nvidia is in early talks to deepen its investment in Reflection AI, in which it has put $800m, or to buy it, the Financial Times reported. One route, an acqui-hire, could avoid a lengthy regulatory review. Neither company had commented when Reuters published.
The maker of the lasers that move data between AI servers says it cannot build enough to meet demand until early 2029. In March the date was the end of 2027. New factories take years, and its customers are now helping to pay for them.
OpenAI has told investors its annualised revenue was approaching $50bn at the end of September, about $20bn below the figure in circulation. The gap comes from how the number is counted rather than from lost sales, but it still knocked the stocks whose order books lean on OpenAI.
GlobalFoundries will make silicon interposers, the slabs that AI processors and their memory sit on, for TSMC at its plant in upstate New York. The $2bn agreement gives GlobalFoundries a part in the AI packaging bottleneck, though volume production does not start ramping until 2028.
Firmus, the Nvidia-backed AI data centre builder, closed the books on its Australian float on Thursday without a clear price, Bloomberg reported. Maas Group, which owns 3.2 per cent of it and supplies it, lost about A$517m of market value instead.
Lambda, the Nvidia-backed GPU cloud, is raising up to $4bn before a planned 2027 listing. Its backlog more than tripled in three months, and one contract signed in late August is worth as much as the whole increase.
SpaceX wants $10bn of bank loans and $30bn of investment-grade debt to buy Nvidia chips, with Apollo leading. Its $25bn bond sale in June is already trading well below par.
DayOne, the Singapore data-centre builder that began inside China's GDS, has filed for a Nasdaq listing. Its prospectus says one short-video company supplies most of its revenue, two of its customers are on the Pentagon's Chinese military list, and members of Congress have already called it China-linked.
Micron will pay Netlist $600m over five years to end every case between them, six days after Netlist asked the ITC to block Micron's high-bandwidth memory and named Nvidia, Broadcom and Google as respondents.
Chinese manufacturers want to sell prefabricated data centres into the American build-out, and CNBC has found one doing it through a Singapore registration. The instrument everybody reaches for does not reach a sale at all: the word 'data center' appears once in the CFIUS regulations, and only at a submarine cable landing. The shortage the Chinese suppliers are selling into is in electrical equipment, not chips.
A projection putting the data-centre AI chip market at $860bn in 2030 has travelled a long way since Sunday morning. It has an author, the research arm of Korea's export credit agency; a policy conclusion it was written to support; and a base year two years stale. Nvidia booked more data-centre revenue in six months than the forecast's entire starting market.
Micron reports on 30 September with a gross margin of 84.6 per cent already on the record and $18bn of customer cash deposits committed behind it. Almost none of the gain came from making more; it came from charging more. Whether a Chinese supply wave can break that is the argument, and the answer is in a filing rather than an interview.
Data-centre securitisations have grown from $4bn outstanding in 2020 to $61bn this year. Follow the chain past the private credit funds and it ends on balance sheets whose liabilities run for decades and whose assets, increasingly, cannot be sold.
The AI build-out is financed in quarters and delivered in half-decades. Berkeley Lab's queue data, GE Vernova's order book and PJM's last capacity auction describe the same bottleneck from three directions, and not one of the three is getting shorter.
Nvidia's outlook for the quarter now ending books no data-centre compute revenue from China at all. That is the company's own sentence, not an inference — and it turns the largest market it is barred from into something closer to a free option than a hole.
A Brookings paper published on Wednesday puts the American AI build-out at $10.3tn by 2032 — larger relative to output than the railroads, electrification or the interstates — and argues its risk has migrated to structures that do not report it. On Thursday, Oracle sent a force majeure notice on a campus whose loans were already trading below par.
The British data-centre developer has disclosed $103bn of contracts ahead of its New York listing. Microsoft and Anthropic account for 85 per cent of it, and $2.6bn is live.
An internal OpenAI presentation projects negative free cash flow of that size through 2030. The number is only shocking until you put it beside the revenue line.
Anthropic and OpenAI have spent a year announcing gigawatts. Both are now hunting 20 to 30 megawatt sites, because the scarce thing is not capital but power that already exists.
Temporal raised $550m at $12.55bn, up from $5bn in February. The number that explains it is not the valuation but the retention: existing customers are spending more than twice what they did.
Semiconductor revenue reached $425bn in three months, a record 31.4% rise on the previous quarter. Memory and price, not volume, did most of the work.
Yields at a level that would normally end a rally have not ended this one. The explanation on offer is that the buying is coming from inside the technology sector.
A final equipment rule tightened authorisation for blacklisted components and did not touch the optical modules that move data inside AI data centres. The market read it immediately.
A forecast taking AI-related chip revenue from $2.6bn to $230bn in six years describes a supply chain, not a product cycle — and the supply has already been bought.
A memory supply agreement is widening toward foundry work, and the number attached to it — 900,000 DRAM wafer starts a month — is the part worth holding on to.
The largest purchasers of AI accelerators are now shipping their own silicon at volume, which changes the negotiation more than the benchmark.
Capital that once chased model training is shifting toward inference, power, and the unglamorous plumbing that keeps deployed systems running.
Features that were uneconomical a year ago are becoming defaults, and product teams are redrawing the line between what runs always and what runs on demand.
Rack densities have pushed past what air can carry away, and the retrofit market is scrambling to catch up with the physics.
Packaging capacity, memory, and power components are the constraints that now shape delivery schedules more than leading-edge logic.
For high-volume tasks with narrow scope, engineering teams increasingly route around frontier models, and the routing layer itself is becoming strategic.
Dozens of specialist providers rented scarce accelerators at premium prices; falling scarcity is now sorting operators from arbitrageurs.
Capable local models are shifting latency-sensitive and private workloads onto consumer hardware, with consequences for chipmakers and cloud bills alike.