OpenAI says 10,000 agents solved a Millennium Prize problem in 88 hours, then the credit fight started
OpenAI announced on September 8 that an unreleased internal model, which it describes as significantly more capable than GPT-6 Astra, worked with roughly 10,000 concurrent agents to produce a solution to the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems the Clay Mathematics Institute listed in 2000. The company published a 166-page proof alongside Lean code that a machine can check step by step. Training on the new model began on August 28. The agent swarm launched on September 1, found the result on September 5 about 88 hours in, and GPT-6 Astra then spent roughly 17 more hours formalizing it in Lean. Across the whole run the agents sent about 4.9 million messages and produced around 300 billion output tokens, with about 2.7 million messages and 130 billion tokens going to Navier-Stokes alone. Before that, a smaller group spent roughly 50 hours on a counterexample for the Euler equation, the inviscid limit, and the resources were then redirected.
The answer is negative. OpenAI says it constructed a vortex that spirals inward along a vertical axis and stretches, thin and long like drying pasta, until the fluid velocity blows up in finite time while kinetic energy stays bounded. Acceleration, pressure gradient, momentum transfer and viscosity have to grow together and cancel each other precisely, which is why the problem resisted about 90 years of attack. Sebastien Bubeck said the final stage of the run cost emphatically millions of dollars, roughly 1,000 times what the company spent on earlier mathematical results. OpenAI says it will not claim the $1 million prize, and the Clay institute's president Martin Bridson noted that evaluation is deliberately unhurried: a solution needs peer-reviewed publication and two years of acceptance in the mathematical community before a committee is even convened.
The dispute arrived before the ink dried. On September 7, NYU mathematician Tristan Buckmaster and Levent Alpöge, who works at Anthropic, published AI-assisted results on three related equations, including Euler with a smooth forcing term. Buckmaster said OpenAI only took up the problem after word of his research spread, and asked whether the Codex sessions where he and Alpöge kept drafts had fed into the model. OpenAI says its researchers never saw the pair's work, that no specific user data was consulted, and that the two proofs differ in approach and in what they establish, but it also said it cannot rule out that de-identified data derived from how people use its products helped improve its models. Buckmaster read that as an admission. Terence Tao weighed in separately, arguing that good open problems are a resource being mined non-renewably, which may be the more durable point of the week.
— OpenAI (official) · AFP / phys.org · Science Net
🔗 OpenAI · AFP via phys.org: OpenAI says 10,000 AI agents cracked one of math's hardest problems in 88 hours · 中国科学报: OpenAI称破解数学"千禧年难题"
Anthropic has reportedly committed up to $517 billion to compute in eleven months
The Information reported on September 6 that Anthropic has signed compute contracts worth up to $517 billion since October 2025, securing at least 14.8 gigawatts of new capacity on top of the one to two gigawatts it already had, and that the company now plans to build its own data centers. Google and AWS account for roughly 11 gigawatts of that, with those two contracts alone estimated at more than $300 billion across ten years. The itemized list runs from an AWS agreement worth $100 billion over ten years and up to 5 gigawatts of Trainium capacity, through roughly $50 billion with Fluidstack and $45 billion with Nscale, to a $9.1 billion, 20-year deal with the bitcoin miner Riot Platforms that converts mining power into AI data center capacity. Riot's stock rose about 25 percent after hours on the news.
For scale, a single gigawatt is roughly one large nuclear reactor's output. In December 2025 Anthropic told investors its server leasing budget through 2029 was about $180 billion. Nine months later the contracted number is nearly three times that. The reversal is sharp because CEO Dario Amodei spent early 2026 warning that competitors were investing too quickly and did not understand the risks they were taking. What changed is Claude: as Claude Code and Cowork took on longer tasks, demand for sustained inference capacity outgrew what the company could buy on demand, and frontier capacity is not something a lab can order when it needs it. Reporting reads the push as locking in a decade of compute ahead of an IPO that is now expected to be marketed no earlier than mid-October.
Two caveats matter. The $517 billion is an upper bound drawn from public contract terms, not an official itemized disclosure or a paid bill; many agreements span a decade and carry adjustment or exit clauses, so what gets used and paid is still open. And Anthropic's own announcements support the direction without confirming the total: in October 2025 it said it would use up to one million Google TPUs, and in November 2025 it committed $50 billion to US data centers with Fluidstack. Bloomberg puts Anthropic's annualized revenue above $65 billion, which is large and still not enough to cover commitments of this size on its own.
— The Information (via The Decoder) · Anthropic (official) · 東方財富
🔗 Anthropic: Expanding our use of Google Cloud TPUs · Anthropic: Investing $50 billion in American AI infrastructure · The Decoder: Anthropic reportedly signs $517 billion in compute deals
Nscale lines up $3.5 billion in pre-IPO money at a $30 billion valuation, with Nvidia writing $2 billion
Nscale, the London-based AI compute provider founded in 2023 and spun out of the crypto miner Arkon Energy, is in advanced talks to raise about $3.5 billion before a US listing that could land as early as September, with Goldman Sachs advising. Nvidia is expected to put in roughly $2 billion as direct equity and Third Point leads about $1.5 billion in convertible notes, with the conversion capped at a $30 billion valuation. That is more than double the $14.6 billion the company reached in a $2 billion Series C in March 2026. The IPO itself could raise a further $3 billion, and the company has about 194,000 Vera Rubin accelerators under contract plus roughly 200,000 GB300 units for Microsoft, across sites in Norway, Texas, Portugal, England and a West Virginia campus due from late 2027.
The driver is one deal. In late August, Nscale signed a six-year, $45 billion compute agreement with Anthropic to supply training and inference capacity on Nvidia's Vera Rubin systems from West Virginia. That single contract pushed Nscale's reported backlog from about $51 billion to $103 billion in a matter of weeks. Those figures need context: Nscale booked $33 million in revenue for all of 2025 and crossed $100 million in a quarter for the first time this year. The $103 billion is signed long-term leases, a standard metric in AI infrastructure where capacity is committed years ahead and cash arrives gradually. It is a projection, not a bank balance.
The structure is the part worth studying. Nvidia funding a company that buys Nvidia hardware has become a recurring pattern in this buildout, and it means the chipmaker is financing demand as well as supply. The valuation also lines up neatly with the rest of the sector: Nebius around $62 billion, CoreWeave around $49 billion, Crusoe at $30 billion on a round that closed September 3, IREN at $18.5 billion, Fluidstack at $18 billion and Lambda at $12 billion or more. Nscale is the one still negotiating its number, which makes the next few weeks the test of whether the market keeps paying this much for contracted megawatts.
— Nscale (official) · Pomegra · Data Center Dynamics
🔗 Nscale · Pomegra: Nscale eyes $30B valuation in $3.5B pre-IPO push · Wortins: Nscale seeks $3.5B pre-IPO raise including $2B from Nvidia
Apple ships its first foldable iPhone, a 2nm A20 Pro and a rebuilt Siri
At its September 9 "Surprise and Shine" event in Cupertino, Apple introduced the iPhone Duo, its first foldable, alongside the iPhone 18 Pro and Pro Max. It was the first keynote under new CEO John Ternus, who took over from Tim Cook earlier in the month and opened by framing the iPhone as an intelligent personal hub. The Duo is a book-style foldable with a 5.4-inch outer display and a 7.6-inch inner screen, 5.02 millimeters thick when open, a titanium frame with ceramic fiber reinforcement, ProMotion, multi-window support and Apple Pencil support over USB-C later in 2026. Apple brought back Touch ID in the power button instead of Face ID. It starts at $1,999 for 256GB, or ¥15,999 in China, with pre-orders on October 16 and shipping on October 23.
The Pro models get the A20 Pro, Apple's first iPhone chip on a 2-nanometer process: a six-core CPU with two new super cores that Apple says run 20 percent faster, four efficiency cores carrying dedicated neural accelerators, a seven-core GPU up to 40 percent faster than the A19 Pro, and a vapor chamber three times the area of the previous generation. The main camera moves to 48 megapixels with a six-blade variable aperture that can open for light or stop down for depth of field. Battery ratings reach 36 hours of video on the Pro and 45 hours on the Pro Max, with wired charging to 50 percent in about 15 minutes. Both phones cost $100 more than last year, and Apple attributed part of the increase to the global memory chip shortage, which is now visible on consumer price tags.
The software story is Siri. The rebuilt Siri AI is launching in beta in English first, with 16 languages planned including Simplified and Traditional Chinese, though it will not be available in mainland China at launch. The first batch of features is modest: spelling correction, point-the-camera-and-ask queries, and photo editing that changes the spatial perspective of a shot. French, Japanese, Korean, Portuguese and Spanish arrive in October. Apple Intelligence runs on device where it can and hands off to Private Cloud Compute when it needs more, which Ternus described as keeping the data inaccessible even to Apple. Alongside the phones Apple announced the Apple Watch Series 12, Watch Ultra 4 with a new readiness score, AirPods 5 with open-ear noise cancellation, and release dates for iOS 27 and watchOS 27.
— Apple (official) · CBS News · HowStuffWorks
🔗 Apple Newsroom · CBS News: The biggest announcements from Apple's tech event · HowStuffWorks: Everything Apple announced today
Memory inventory at Samsung and SK Hynix drops below ten days as HBM4 eats wafer capacity
A KB Securities report dated September 7 found that the sellable memory inventory at Samsung Electronics and SK Hynix has fallen below ten days of supply, against the four to six weeks that manufacturers normally hold to absorb demand swings and line maintenance. The same report raised its forecast for hyperscaler AI infrastructure spending in 2027 to $1.3 trillion, up about 60 percent year over year, and projected that memory's share of that spend climbs from 14 percent in 2025 to 40 percent this year and 57 percent in 2027. TrendForce is more aggressive, putting DRAM and NAND together at as much as 68 percent of major cloud capital expenditure in 2027.
The mechanism is not simply that AI needs more memory. HBM4, the next-generation high-bandwidth memory built for training and inference, consumes roughly three times the wafer capacity of conventional DRAM per unit of capacity. With total wafer supply roughly fixed, every shift of advanced capacity toward HBM4 reduces the bits of standard DDR5 and LPDDR5X that the same wafers can produce, which pushes the shortage from one product line into the whole server and consumer stack. KB Securities expects 2027 bit demand growth to outrun supply growth by more than ten percentage points for both DRAM and NAND, and its research head Kim Dong-won described the market as facing the possibility of running out of sellable inventory.
Samsung and SK Hynix together hold more than 70 percent of global DRAM and NAND, so their inventory position is the whole industry's problem. Meaningful new supply is not expected to arrive before 2028, and the pricing power that comes with scarcity is already showing up in contract negotiations. The market has been slow to price it: both stocks are down more than 30 percent from their peaks over the past three months and trade at roughly three times next year's estimated earnings, which suggests investors are treating the shortage as a margin risk for buyers rather than a windfall for suppliers.
— KB Securities (research report) · TrendForce · BusinessKorea
🔗 TrendForce · 电子工程专辑: 存储芯片库存见底，三星、SK海力士供货量已不足10天 · 新浪财经: 涨不停！三星SK海力士库存不足10天
Physical AI moves onto the factory floor: Tesla's Optimus plant rises while China's lines start running
Drone footage captured on September 9 shows steel assembly accelerating at Tesla's dedicated Optimus factory on the North Campus of Gigafactory Texas, with the mid and south bays filling in over the footing field and the structure pushing toward the north grade beam. The project broke ground on March 23, 2026, showed its first steel beams in late May and was about 25 percent through its superstructure by mid-August. The planned footprint is 5.2 million square feet and more than 4,000 feet long, close to the length of the existing main Giga Texas building, and it sits next to the planned Terafab AI chip factory so that the robot's processor and its body come out of the same complex. Tesla is targeting shell completion by the end of 2026, high-volume production in summer 2027, and a long-term annual capacity of 10 million robots.
Analysts tracking the ramp put near-term numbers far below that ceiling. Nomura raised its estimate for the Fremont Optimus Gen 3 line from about 50,000 units a year to roughly 70,000, with another 70,000 planned in Austin in 2028, and expects around 25,000 Optimus shipments in 2026 with a September weekly target that could reach 1,000 units. Fremont's line started in mid-2026 mainly to supply Tesla's own operations. The gap between a 10-million-unit ambition and a 25,000-unit year is the honest measure of where humanoid manufacturing actually stands.
China is running the same experiment in parallel and with more shipping evidence. XPeng switched on its IRON production line on September 8 and the first unit walked off unaided, which the company calls the world's first automated line for high-end general-purpose humanoids. At the Hangzhou AI and Robotics Expo that opened September 9, Unitree robots were already doing hubcap pickup, transport and installation work at a Geely plant, and a Zhejiang innovation center said 2,000 clothing-scenario robots are entering batch delivery with 99 percent success in specific test scenarios. The United States moved the other way on supply, with the FCC banning new imports of foreign-made humanoid and quadruped robots on national security grounds. Production lines, takt time and yield are replacing demo videos as the metrics that matter.
— Tesla (official) · Joe Tegtmeyer drone footage · 机器人大讲堂
🔗 Tesla · BASENOR: Giga Texas Optimus factory steel work surges in latest drone footage · 机器人大讲堂: 人形机器人"进厂打工"，浙江找到了一条从展台到产线的通路
OpenAI and DeepSeek cut prices in the same week as the model market turns into a cost fight
OpenAI CFO Sarah Friar used a Goldman Sachs conference in San Francisco on September 7 to lay out a pricing shift. She said OpenAI is targeting chip design, life sciences and financial services, is experimenting with charging for business outcomes rather than raw consumption, and recently cut the price of its low-cost Luna model by 80 percent, which drove roughly a tenfold increase in usage. Codex, the company's coding tool, now has 25 million users. Enterprise revenue rose 32 percent from June to July against 20 percent growth in overall annualized revenue, and the enterprise and consumer halves of the business drew even by mid-year, ahead of the year-end target. Friar also argued that running Luna through a cloud provider is cheaper than deploying Z.ai's GLM 5.3 the same way, which reframes open weights as a total-cost question rather than a sticker-price one. On chip design she cited OpenAI's Jalapeño inference chip, developed with Broadcom, which reached tape-out in under nine months using OpenAI's own models, with early samples showing about 50 percent cost savings against typical AI GPUs.
DeepSeek answered from the other direction. On September 9 it announced cuts to its Flash-series API pricing effective September 10 at noon Beijing time, with cache-hit input falling from 0.05 to 0.02 RMB per million tokens, a 60 percent reduction, cache-miss input from 1.5 to 1.0 and output from 4.5 to 4.0. Peak-hour pricing remains double the off-peak rate. The cut lands less than a month after DeepSeek raised V4 Pro prices by as much as 11 times, and it does not fully restore the old schedule: output at 4 RMB is still double the 2 RMB that applied before August. The steepest reduction is on cache-hit input, the line item that matters most for RAG pipelines, multi-turn agent workflows and code completion, where repeated context is reused across calls.
The two moves together describe where the market has gone. OpenAI is cutting its floor price to defend against open-weight alternatives while testing whether buyers will pay for measured outcomes instead of tokens, and DeepSeek is compressing the cost of the workloads that agent products depend on while keeping a margin buffer on generation. DeepSeek also began internal testing of an intermediate V4.1 Flash version on September 8, with a new architecture, native multimodality and improvements claimed on capability, speed and cost. For anyone building on these APIs, the practical question is no longer the price per million tokens but the cost of finishing a task, which includes retries, tool calls and human review.
— OpenAI (official) · DeepSeek (official) · Reuters · 澎湃新闻
🔗 OpenAI · Reuters via The Star: OpenAI offers AI for chip design, touts cost advantage over open-source · Quartz: OpenAI CFO says Luna undercuts Chinese AI on price · 澎湃新闻: 刚涨完价，DeepSeek又"砍价"六成
Next digest: September 11, 2026.
Top comments (0)