AI News

Edge AI Daily Morning Brief: Bernstein estimates OpenAI's compute spending could reach $614 billion…

TMTPost · AGIUpdated Nov 9 12:26

AI summary

In a research report dated November 8, Bernstein estimates that OpenAI's disclosed compute agreements imply actual infrastructure spending of about $614 billion through 2030, below the $750 billion cited by the market, with peak spending pushed back to about $184 billion in 2031. Anthropic agreed to take over OnePWR Solutions' order for 1.7 gigawatts of Caterpillar heavy-duty gas generators for $3.4 billion, bypassing the grid interconnection queue. Political resistance to data centers has risen following the U.S. midterm elections, and the newly elected governor of Texas has pledged to pause data center construction statewide.

Body · AI translation

Silicon Valley Frontier:

I. Political Pressure and Debt Costs Compound: Data Center CFOs Brace for a Difficult Road Ahead

1. Political resistance to data centers surges after the U.S. midterm elections: Gina Hinojosa, the newly elected governor of Texas, has pledged to suspend data center construction statewide upon taking office until the state legislature passes regulations; about 40% of planned capacity nationwide is slated for states that advocate restricting construction, with Texas alone accounting for roughly 25% of U.S. planned capacity through 2030 and carrying the highest risk exposure, and policy clarity may not arrive until late May 2027.

2. Political risk quickly transmits to the market: on November 5, the Nasdaq Composite fell 5.5% in a single day and entered a technical correction, the Philadelphia Semiconductor Index plunged more than 12%, and Nvidia's market value evaporated by more than $400 billion in one day; credit markets turned bearish in tandem, with CCC spreads soaring from about 904 basis points in early May to about 1,267 basis points, a new high since 2023, and CoreWeave's B+ rating and leveraged loan yields above 10% signal that speculative-grade AI infrastructure borrowers are being repriced.

3. The AI data center financing chain is fracturing along tiered lines, with tenant credit becoming the core of pricing: Google-backed Fluidstack carries a spread of about 195 basis points and Amazon-leased deals about 150 basis points, while CoreWeave-leased bonds carry about 400 basis points; AA-rated tech giants' financing costs are about 5% to 5.5%, and Oracle, at the lowest rung of investment grade, would trigger forced institutional selling if downgraded. CFOs must manage three statements at once—political risk, financing stratification, and cash flow resilience—and industry pricing power is shifting from "who has projects" to "who has cheap capital and reliable permits."

1. Elections and referendums: data centers became a core issue in the U.S. midterm elections on November 3, 2026; according to Ballotpedia, there were 50 related measures across 13 states during the year, and of the 47 decided as of November 4, 46 supported restricting development, with the opposition camp spanning party lines (both red and blue states), and gubernatorial races in Texas, Ohio, Georgia and other states all running on data center regulation platforms.

2. Market reaction: on November 4 the market diverged sharply, with power stock Vistra falling 13.8% in a single day (KeyBanc downgraded it the same day), while hyperscale cloud providers rose against the trend (Amazon up 8%, Google over 6%, Microsoft up 4%) and CoreWeave fell 10%; the pricing logic shifted to "who can withstand delays," favoring incumbents with ready capacity and solid balance sheets.

3. Trend judgment: AI infrastructure has entered the "permitting era," with permits becoming the third major constraint after capital and chips; UBS estimates that about 30% of planned U.S. capacity in 2028 faces construction suspension risk, with the impact being slower development and higher capital costs rather than stalled demand, and development will concentrate among participants able to absorb delays.

1. The U.S. AI rally slumps: on November 5 the Dow fell 1,833 points (-3.44%) to 51,523.06, the Nasdaq fell 5.52% to 24,315, and the S&P 500 fell 4.36% to 7,353.73, the worst single day since April 2025; Asian markets followed the next day, with South Korea's KOSPI briefly falling below 5,000 intraday and dropping 13.19% for the week, Taiwan's weighted index closing at 39,701.64 and falling about 12% over the week into a technical bear market, and the Nikkei 225 dropping more than 6%, as semiconductor stocks such as SoftBank and Tokyo Electron were hammered.

2. The plunge stems from three compounding factors: Democrats won a landslide in the midterms and captured states with concentrated data center plans, with most of the 41 measures to restrict data centers passing, raising uncertainty over AI infrastructure approvals and regulation; free cash flow at Meta, Amazon and Alphabet turned negative, casting doubt on AI capital expenditure returns; and positioning was overly crowded, with a 3x leveraged long semiconductor ETF falling 39% in a single day, the VIX rising to 28.18, and the spread between spot and three-month futures widening to its largest since April 2025, signaling deeper declines ahead.

3. The key inflection point is Nvidia's fiscal 2027 third-quarter earnings release after the U.S. market close on November 17 Eastern Time: the company guided revenue of $108 billion (plus or minus 2%) and gross margin of 74.0% (plus or minus 50 basis points), excluding data center revenue from mainland China and Hong Kong; an upside surprise could drive a rebound in unjustly sold-off AI supply chain stocks, while a gross margin miss below 74% could see valuation cuts spread from semiconductors to the entire tech sector.

IV. Anthropic Spends $3.4 Billion on Gas Generators as AI Labs Race for Their Own Power

1. Procurement scale: in November 2026, Anthropic agreed to take over a 1.7-gigawatt Caterpillar heavy-duty gas generator order from OnePWR Solutions for $3.4 billion (about 99% of Caterpillar's $3.421 billion in power generation equipment sales in the third quarter of 2026), the largest single power equipment purchase on record, essentially buying out priority to bypass grid queues.

2. Financial gap: Anthropic's updated IPO filing discloses nearly $900 billion in compute and infrastructure commitments through 2045, of which about $448 billion must be paid before 2030 and about $70 billion comes due before September 2027; meanwhile third-quarter revenue was $17 billion (up more than 12-fold year over year) and net loss was about $23.3 billion, a stark mismatch between revenue and commitments.

3. Industry logic: Grid interconnection queues in major U.S. markets commonly run 5 to 7 years; Wood Mackenzie estimates gas turbine prices will rise to $600 per kilowatt by the end of 2027 (up 195% from 2019), while global orders of 110 gigawatts by the end of 2025 far exceed production capacity of 60 to 70 gigawatts. Electricity is replacing chips as the new moat in AI competition.

V. World Models Are About to Be Everywhere: Five Things You Need to Know

1. Background and asset layer: In March 2025, Niantic announced the sale of its gaming business, including Pokémon Go, to Scopely for $3.5 billion (closing in May of that year), while retaining the roughly 30 billion image scans accumulated by players; it spun off Niantic Spatial to pivot to world models, using real-world environmental data to reconstruct the digital world.

2. Capital and deal layer: In 2026, capital poured into world models—AMI, founded by Yann LeCun, completed a $1.03 billion seed round in March of that year at a valuation of about $3.5 billion; on September 28, AMD announced an all-stock acquisition of Fei-Fei Li's World Labs for about $8.2 billion (its second-largest acquisition ever), with Fei-Fei Li becoming AMD's chief scientist.

3. Technology and landscape layer: World models center on predicting the "next visual state" and understanding physics and causality, distinguishing them from language models that predict the next word; their data sources are coming from ordinary people's daily lives (players scanning street views, gig workers recording household chore videos); the moat lies in real-world data and compute chips. Around 2027, home robots and autonomous driving may accelerate toward deployment, but they remain constrained by data quality, compute costs and generalization capability.

VI. No Longer Just Footprints on the Moon: NASA Pushes Lunar Base Toward Bidding With 8 Companies and a Nuclear-Powered Lunar Rover

1. On November 7, 2026, NASA unveiled the CLPS 2.0 framework at the MAX POWER aerospace exhibition at Kennedy Space Center in Florida, selecting 8 companies as finalists. The framework is a 10-year, multiple-award indefinite-delivery/indefinite-quantity (IDIQ) master contract with a total ceiling of $10 billion; at the same time, the CLPS 1.0 contract ceiling was raised from $2.6 billion to $4.2 billion, with the original contract running through November 2028 and awarded missions extendable to November 2033. Core change: payloads upgraded from "scientific instruments" to "infrastructure."

2. In the same announcement, NASA established the nuclear-powered lunar rover PROMISE (polar roaming, mapping and in-situ exploration), a six-wheeled vehicle to land at the lunar south pole, capable of operating during roughly 14-day lunar nights and in permanently shadowed regions, while advancing supporting elements such as fission surface power systems and lunar surface communications/solar towers. Background logic: permanently shadowed craters at the south pole hold water ice (which can produce oxygen and hydrogen), but harsh lighting renders solar power ineffective, so continuous power must be provided by radioisotope power sources or fission reactors.

3. The 8 finalist companies fall into three categories: pure commercial lunar logistics (Intuitive Machines, Firefly), heavy-lift launch and lander giants (SpaceX, Blue Origin), and traditional defense contractors (Lockheed Martin, Northrop Grumman). CLPS is an IDIQ framework; being selected does not mean winning orders, which will concentrate among a small number of the most reliable players; defense giants stand out in systems integration and orbital transportation.

VII. Third Proposal Falls in Five Months: Stockland Abandons Sydney Data Center Due to Community Opposition

1. Phenomenon and background: Veteran Australian property developer Stockland and joint-venture partner John Boyd Properties withdrew from an AI data center plan in the Cooks River area of southern Sydney, on an 18.3-hectare site at the former Kogarah Golf Club in Arncliffe, originally part of a A$3.5 billion logistics hub; this is the third Sydney data center proposal canceled due to community opposition in five months, after Goodman withdrew Project Mars in Lane Cove (about A$1.2 billion, 90 megawatts, only about 20 meters from the nearest homes) at the end of September.

2. Policy and trends: The New South Wales government released a data center policy framework on August 17, 2026, trading "fast-track approval" (assessment within 2 months, state government stage no more than 75 days) for developers bearing additional energy and water costs—they must sign renewable energy agreements of no less than 10 years (at least 40% from wind power), have storage of no less than 25% of generation capacity lasting 4 hours, and fully offset temporary potable water use. The background is a double squeeze of supply and demand: AEMO forecasts that data centers' share of NSW grid supply will rise from 5% in 2026 to 11% in 2030, while about 7 gigawatts of coal-fired units are scheduled to retire over the same period.

3. Corporate impact: Analysts say the new framework turns industry access into a capital screening process, eliminating small developers and speculative proposals (Oxford Economics estimates that of every 7 megawatts of grid connection applications, about 6 megawatts are "phantom demand"). Although Stockland withdrew from Cooks River (interpreted as a tactical stop-loss), it still retains 3 projects in the NSW Investment Delivery Authority's fast-track approval queue (Wetherill Park, Kemps Creek, Macquarie Park), and has formed a 50:50 joint venture with EdgeConneX to develop hyperscale campuses in Sydney and Melbourne, having already secured 450 megawatts of power; its share price closed at A$4.60, with a market capitalization of about A$11.19 billion.

VIII. Bernstein: OpenAI compute partnership spending could reach $614 billion

1. A research report led by Bernstein analyst Mark Moerdler (November 8, 2026) estimates: after breaking down OpenAI's disclosed compute agreements deal by deal under a bottom-up model and converting them according to the pace of phased data center commissioning, actual infrastructure spending before 2030 will be about $614 billion (about RMB 4.1 trillion), significantly lower than the widely cited market figure of $750 billion; the spending peak is postponed to about $184 billion in 2031, and the report states bluntly that the premise that "all contracts will be delivered on time" is "highly unlikely" to hold, and the actual amount may shrink further due to construction delays and project cancellations.

2. Signed amounts do not equal actual spending: first, chips from Nvidia, Broadcom, AMD and others are mostly financed by the suppliers themselves and paid in installments over a hardware useful life of about 6 years; second, the agreements between OpenAI and AMD and Cerebras embed stock-price-linked warrants, which can bring the cost of covering part of the compute to close to zero; third, Georgia's Project Camellia (about $20 billion, 3.2GW) is treated as an upfront cost, amplifying early book pressure. Contracts with Oracle, CoreWeave and others mostly state "caps" rather than committed revenue.

3. Spending structure diverges: the share of hyperscale cloud providers will fall from about 82% in 2026 to a peak of about 64% in 2031 and about 15% in 2035, with higher near-term revenue certainty; specialized compute suppliers will face pressure on renewals and bargaining power after supply and demand ease. Bernstein maintains "outperform" ratings on Microsoft and Oracle (target prices $722 and $325), and maintains an "underperform" rating on CoreWeave (target price $54). Over the same period, OpenAI's annualized revenue as of September 2026 was about $50 billion (up about 77% year on year), and peak spending in 2031 is about 3.7 times that, with the gap relying on financing to fill (it completed $110 billion in financing in 2026, with Amazon $50 billion, Nvidia $30 billion, and SoftBank $30 billion).

IX. Apple saves $5.5 billion in taxes in Singapore, but the global minimum tax is closing in

1. Apple's fiscal 2026 annual report shows: tax incentives and relief from Singapore reached $5.5 billion (the largest overseas item), but Singapore simultaneously levied a Qualified Domestic Minimum Top-up Tax (QDMTT), taking back about $2.5 billion, cutting the net benefit by nearly half and directly pushing up the company's effective tax rate.

2. Tax rate changes: Apple's effective tax rate rose to 17.0% in fiscal 2026, higher than 15.6% in fiscal 2025; among this, the Singapore tax rate differential contributed a 3.4 percentage point reduction, while the QDMTT top-up tax added back 1.6 percentage points, reflecting the first report card after the OECD "Pillar Two" global minimum tax took effect.

3. Rules and trends: Pillar Two sets a 15% tax rate floor by jurisdiction, and Singapore, through the Multinational Enterprise (Minimum Tax) Act 2024, taxes low-taxed domestic profits first, turning a "tax haven" into a "tax toll booth"; as of October 2025, more than 65 countries have implemented the agreement, at least 60 countries have completed domestic legislation, and at least 22 of the EU's 27 member states have incorporated it into domestic law. Tax competition among multinational enterprises has shifted from "profits flowing to low-tax locations" to "pressing down actual tax rates in each jurisdiction."

X. Completion rate only 65.8%: NHTSA bets on OTA remote fixes, recall system faces biggest reform

1. Background and pain points: At the NHTSA "Recalls Next" summit on November 5, 2026, it was disclosed that in 2025 the agency oversaw nearly 900 vehicle safety recalls involving more than 29 million vehicles; its report shows that from 2012 to 2022 the weighted average recall completion rate was only 65.8%, and about one-third of recalled vehicles continued on the road with defects. About 40% of owners cited "inconvenience" as the main reason for not completing recalls; the completion rate was 68% for new cars (three years old) and only 18% for old cars (fifteen years old).

2. Reform direction: Administrator Morrison is pushing recalls centered on OTA remote software updates, turning the repair action from "an owner action" to "a vehicle action" (for example, Tesla's December 23, 2023 23V-838 recall of about 2 million vehicles was completed through a single OTA update); three supporting measures—improving notification methods, allocating $1.5 million to states to embed recalls into vehicle registration renewal notices, and launching a bulk VIN lookup tool (up to 100 VINs at a time).

3. Boundaries and impact: Completion rates drop sharply with vehicle age (about 90% for under three years, only 38% for sixteen to twenty years), while the average age of vehicles in the United States has reached 12.8 years, making it hard for old cars to enjoy the OTA dividend; regulators emphasize that OTA recalls and traditional recalls apply the same standards, and the public comment window is open until December 5, 2026. In the long run, recall completion rates may be included in public ratings, affecting brand reputation and used-car residual values.

XI. 32% of British children use it, and 40% of child sexual communication crimes point to Snapchat—it has finally closed "friend recommendations"

1. Regulatory action: In November 2026, the UK Office of Communications (Ofcom) confirmed that Snapchat has turned off the "find friends" feature by default for users under 18 in the UK, requiring full deployment within this month and enhanced age verification by the end of the year; under the Online Safety Act 2023, violations can be fined up to £18 million or 10% of global revenue.

2. Risk data: Ofcom data from May this year showed that 32% of British children aged 8 to 12 use Snapchat, while its legal minimum age is 13; an NSPCC analysis of police data in 2025 said that 40% of child sexual communication crimes on identified platforms occurred on Snapchat, and the failure of age verification made protective design virtually useless.

3. Industry and market impact: The case establishes a new compliance benchmark that "features involving child contact risk should be off by default," forcing a reversal of the logic in which social products drive growth by expanding connections; for businesses operating in the UK and Chinese overseas apps (TikTok, Shein, Temu, etc.), it creates cumulative compliance thresholds for default design, age verification, and algorithmic risk assessment.

1. October 2026 was called "Black October" for cybersecurity by some practitioners: high-risk vulnerabilities were exposed one after another in foundational software such as PHP and Node, and attacks took on a new "chain-like" form—internal credentials at AI inference service provider Fireworks leaked, affecting the AI code editor Cursor, whose malicious plugin was distributed to as many as about 5,000 enterprises; on October 29, about $6.3 billion in assets at Binance were stolen in about 20 minutes, with law enforcement pointing to a North Korea-linked hacking group, and Elliptic said it was "almost certainly the largest single theft known to date."

2. AI reverses the cyberattack "kill chain": open-source tool ARTEX connected to large models such as GLM-5.3 can discover and exploit vulnerabilities in real time, and the attack threshold has plummeted; Reuters disclosed on October 19 that hackers spent only about $5,000 in tokens, at a cost of about $50 per attack, to breach more than 100 companies and over 1,200 e-commerce sites in the US, Canada, Australia and Japan; a CrowdStrike report said similar attacks from late September to early October caused data leaks at 7 South Korean financial institutions and about 68,000 people.

3. Former Facebook security chief and current Cognition chief security officer Stamos sharply criticized the cybersecurity industry, worth about $250 billion to $300 billion, as having "long deserved a reckoning": defense logic is static while attacks are dynamic, Microsoft's "Patch Tuesday" patch count has risen from 150 to 200 to more than 900, and supply chain attacks render single-point defenses ineffective; CISOs, because their authority and responsibilities do not match, have become "the first lambs to be slaughtered." He judges that the industry will first go through a "dark age" lasting several years, with the turning signal being a decline in the number of major vulnerabilities, worm-level vulnerabilities and patches.

Source

TMTPost · AGI