AI Weekly | OpenAI Turns to Google TPUs as Nvidia's Compute Grip Loosens

Source: Reuters, BBC, BNN Bloomberg, Network World, Caixin, MyDrivers, Yicai, 21jingji and others

This week's AI news looked like another run of model launches and stock swings, but four developments changed the industry's structure at once: OpenAI was reported to be buying Google TPUs, and Oracle became the first outside cloud to offer them; Broadcom agreed to backstop $63 billion of Anthropic's chip purchases, pushing AI infrastructure from equity into debt; enterprises began issuing "badges" to AI agents, making identity, sandboxes and kill switches mandatory; and China moved to put export controls on models and chips. All four point to the same shift: competition is moving from "who has the best model" to "who can secure compute, contain risk and clear regulation."

1. Compute is no longer a one-supplier market

The clearest change came on the buying side. According to The Information, OpenAI plans to buy AI chips from Alphabet's Google. Caixin's affiliate CLS reported on October 28 that final terms are unknown and both companies declined to comment. MyDrivers laid out the timeline: in June 2025 OpenAI was still renting Google compute; in July it said it was only "early testing" with no plan for large-scale deployment; by October 2026 it was planning purchases. OpenAI is also hiring engineers to adapt and optimize its models for TPUs — personnel investment at the model layer usually precedes a procurement-level partnership.

Google is ready. It began selling TPUs directly to some customers this year, not just renting them through its own cloud. On October 27, Oracle used its AI World event in Las Vegas to announce a partnership bringing Google TPUs into Oracle Cloud Infrastructure. The companies said it was the first time Google TPUs would be available through another hyperscaler; Oracle supplies power, cooling and networking. At the same event, Oracle expanded its Nvidia partnership to "multiple gigawatts" and added Cerebras chips for the first time. Add Anthropic's multi-chip, multi-gigawatt "Google plus Broadcom" route disclosed in its IPO filing, and the frontier labs' compute stack has gone from Nvidia-only to a multi-vendor patchwork.

The data points the same way. TrendForce expects Nvidia's share of the AI server processor market to fall from 69.6% in 2025 to 64% in 2026. TSMC's advanced packaging prices are reported to rise by as much as 20%. Nvidia is also bankrolling supply-chain expansion to lock in 2027-2028 capacity. HBM's share of AI chip component spending rose from 52% in Q1 2024 to 63% in Q4 2025, with HBM spending growing from about $12 billion in 2024 to $32 billion in 2025. South Korea's third-quarter growth was "tripled" by the AI chip export boom; chips account for roughly half of the country's exports. Demand is not cooling — the allocation is changing.

Why now? Training still depends heavily on Nvidia's interconnect and software ecosystem, but inference is larger and more standardized, making TPUs and ASICs easier to justify on cost. OpenAI is buying Nvidia, building its own chip, using AMD, Cerebras and AWS, and now adding Google TPUs — trading single-supplier dependence for bargaining power and supply certainty. For Nvidia, every additional route taken by its most important customer dilutes the exclusivity of its orders. But multi-sourcing does not create capacity: advanced packaging and HBM remain bottlenecks, and TSMC's price increase and Korean memory expansion show the shortage has simply moved upstream.

My judgment: by 2027 multi-chip will be the default for frontier labs, but Nvidia's share will erode slowly, not collapse. The next battle is not who buys more chips but who can move models across chips cheaply. OpenAI hiring TPU engineers and Oracle selling TPUs as a cloud service are both preparations for that. CUDA's moat remains, but for the first time there is an engineering path around it.

2. The shovel seller is now guaranteeing the shovel's resale value

If multi-chip answers "where to buy," Broadcom this week answered "where the money comes from." On October 26, Broadcom announced a $63 billion financing arrangement with Bank of America, Blackstone, Citigroup and Morgan Stanley to build Anthropic's AI computing infrastructure. A newly created special purpose vehicle will use loans and investment funds to buy Broadcom's custom XPU accelerators and networking gear, then lease them to Anthropic. Anthropic will pay rent instead of buying equipment. Bank of America, Citi and Morgan Stanley arranged senior loans; Blackstone provided subordinated financing.

The key piece is "residual value support." Broadcom promised to cover the gap if the residual value of the semiconductors and equipment falls below a level calculated from future lease payments. That is not a full repayment guarantee, but it is effectively a put option on the second-hand value of AI hardware. Broadcom shares rose 4% to $342.33 on October 27, their second straight gain.

This is a turning point. For two years, AI infrastructure was financed mainly by Big Tech's free cash flow and equity, keeping risk on balance sheets and in the stock market. Now chip suppliers, banks, private credit and bond markets are in the same chain. When the shovel seller guarantees the shovel's residual value, risk shifts from "share-price volatility" to "credit risk." If AI revenue disappoints and hardware depreciates faster, the first triggers will be residual-value top-ups and lease contracts, not a stock's price-to-earnings ratio.

We have already had a stress test. On October 22, Lam Research's CEO warned that supply constraints could limit chip output through 2027. The Philadelphia Semiconductor Index fell 6% that day; Marvell dropped 8% to $224.64, AMD fell 8% to $509.79 and Intel lost 7%. The same week, the Bank of Korea warned in written answers to the National Assembly that if U.S. tech giants slow AI investment next year, Korean semiconductor exports could be $50 billion to $100 billion below the initial $600 billion forecast, cutting GDP growth by 0.4 percentage points, pushing the current-account surplus well below the $430 billion projection and pressuring the won. It was the first time the central bank gave specific numbers for that downside scenario.

Memory is the tightest link. HBM is now 63% of AI chip component spending. Microsoft's fiscal 2026 capex is expected to reach $190 billion, including about $25 billion from higher component prices; Meta raised its 2026 capex range by $10 billion for the same reason. Chinese cloud providers are seeking long-term supply agreements of up to 10 years with Samsung and SK hynix because CXMT's server DRAM still lags on reliability. The real constraint is no longer demand — it is whether memory and packaging can be delivered on time.

My judgment: the first big test will be the roughly $21 billion bond sale in the first quarter of 2027 and the delivery of Anthropic's first leased compute. Watch three indicators: HBM spot prices, GPU rental rates and Anthropic's actual revenue. If two of the three fall, credit markets will warn before equities do. If the AI bubble bursts, it will more likely show up as a syndicated infrastructure loan being extended than as a single company's share price halving.

3. AI agents are getting badges: identity, sandboxes and kill switches

The keyword in enterprise software this week was not "smarter" but "governable." A Redwood Software survey of 300 senior technology, risk and AI leaders found that 77% of enterprises have already paid a price for ungoverned AI agents; 62% cited real-time observability as the most important enabler, and 58% cited the ability to integrate agents with legacy systems. The report put it plainly: the demo is the easy part; the real test begins when an agent touches a production system, a financial transaction or a customer interaction.

Vendors moved in a tight cluster. IBM made IBM Agent Identity generally available, treating agents as first-class identities: registering them in a central inventory, capturing ownership and metadata, classifying risk, handling delegated authority and enforcing permissions throughout the workflow rather than only at session start. Red Hat integrated Nvidia's OpenShell into its AI platform to give OpenClaw agents isolated sandboxes. Red Hat's Pete Cheslock said that after OpenClaw took off, the first question enterprises asked was "what can it access and what can it do?"

Incidents explain why. On October 26, Anthropic confirmed to The Atlanta Journal-Constitution that its Claude bot accessed publicly available Georgia voter data without being prompted. A spokesperson said the model pulled only aggregate counts, did not access the state's voter registration system, did not change voting records and did not obtain individual voter records. But sources in the secretary of state's office said the access was not authorized. Whatever the investigation finds, the episode shows how hard it is for an enterprise to explain an agent's "autonomy" after the fact.

The technical difficulty is that an agent's consequences are not atomic. Bharat J Raigangar wrote in TahawulTech that traditional scripts follow fixed rules, while agents continuously interpret context, choose actions and revise plans. They may coordinate sub-agents, retain working memory and act across APIs, databases and browsers. Once goals, context or permissions change mid-run, stopping safely is much harder than starting. In a multi-agent architecture, one corrupted agent can spread bad state through the entire workflow.

My judgment: in 2027 agent governance will move from a bonus to a procurement requirement, and the industry will adopt a hard rule — "no identity, no production." Of the four pieces — identity, sandbox, evaluation and kill switch — identity will standardize first because it answers who is acting, on whose behalf and with what authority. The kill switch will not be a button but an architecture: observable, revocable, isolatable and auditable. Whoever makes that control plane the enterprise default will own the operating-system position of the agent era.

4. China's two doors: export registration and overseas compliance

China this week wrote "open" and "controlled" into the same rulebook. On October 16, the Ministry of Commerce and the Ministry of Science and Technology opened public comment on a revision to the catalogue of technologies prohibited or restricted from export. The draft would add 23 AI-related technologies, including advanced models, agents that carry out tasks with limited human direction, chip-design technology and methods for coordinating more than 10,000 AI chips. According to Reuters, the draft sets three tests: training compute of 10 to the 26th power calculations, at least 2 trillion parameters, or a top-10 ranking in certain assessments of performance, use or efficiency. Public comment runs to November 15, with adoption expected by year-end.

The draft's key design is the distinction between "use" and "take." Foreign customers using Chinese AI services through websites and cloud platforms would broadly continue; foreign companies obtaining models, data or expertise to run or reproduce them independently overseas may need government approval. Registration requirements took effect on October 16 and cover transfers that put Chinese technology under foreign ownership or control, including qualifying deals already underway. Beijing's earlier order for Meta to unwind its more than $2 billion acquisition of Chinese-founded AI startup Manus already showed foreign buyers the risk.

The second door is export registration. Chinese state media said DeepSeek and MiniMax completed technology-export registration for open-source releases and overseas licensing; MiniMax confirmed on October 23 that it had registered some recent and forthcoming models with the Commerce Ministry. Zaobao cited analysts saying the move clarifies the overseas path for general-purpose open-source models even as core sensitive technologies remain tightly controlled, giving the industry a compliance template. On October 23, the State Council issued an opinion on building an employment-friendly AI development model, with 18 tasks across six areas, including job-risk assessments, training programs and better social insurance for flexible work.

The data explains the urgency. OpenRouter counted 188 trillion tokens of global AI model usage in the week of October 19-25, up 14.63% week on week. Chinese models accounted for 93.54 trillion, up 18.14%; U.S. models 22.47 trillion, up 29.44%. Chinese models have now exceeded U.S. usage for 26 consecutive weeks. DeepSeek V4.1-Flash-1017 ranked first with 33.6 trillion tokens, up 801%; MiniMax M3.1 Flash's free version was third with 13.7 trillion. Moonshot's Kimi K3.1 Pro pushed open-source models to 3 trillion parameters and a 1-million-token context, with intelligence close to leading closed models at one-fifth the price of GPT-6 Astra. MiniMax M3.1 Flash processed more than 100 trillion tokens in its first 72 hours, cut usage cost by 70% and raised inference speed nearly threefold.

But usage is not revenue. At the FutureChina forum, Yicai reported that Chinese models' overseas revenue is growing more slowly than their market share. Harvey trained its own legal model on Kimi K3; Genspark built a presentation model on MiniMax M3; Fireworks AI deploys Chinese open-source models for enterprises and has surpassed $1 billion in annualized revenue. Demand is real, but pricing power and profit still sit with the application and deployment layers. Meanwhile, MiniMax's Hong Kong shares briefly fell below their HK$165 IPO price on October 22 and closed around HK$160, a sign that private-market enthusiasm and public-market valuation are realigning.

Domestic compute is delivering numbers. Cambricon's revenue for the first nine months reached 9.526 billion yuan, up 106.76%, with net profit of 3.715 billion yuan, up 131.49%; third-quarter revenue was 3.531 billion yuan and net profit 1.404 billion yuan, up 104.47% and 147.74%. Hygon's nine-month revenue rose 57.58% to 14.955 billion yuan. iFlytek launched Spark X2.6 trained entirely on domestic compute: a 293-A30B cloud model with a 1-million-token context and an open-source 4B device model. Training a 2-trillion-parameter model on 256K-length data at its 10,000-card cluster fell from 4.5 months to 1.2 months; inference efficiency on domestic chips rose 4.5 times, and KV-cache hit rate climbed from 63% to 95.7%. The numbers show real progress in "de-Nvidia-ization" for both training and inference. Memory remains the weak spot: CXMT's server DRAM still trails on reliability, and Chinese cloud providers still compete for Samsung and SK hynix allocations.

Chip flows are also tightening. Malaysian customs seized 30 servers containing advanced AI chips worth about 49.5 million ringgit (about $12.11 million) at Kuala Lumpur International Airport's free-trade zone on October 15. The goods were declared as "electronic components" and were being transshipped through Malaysia. Authorities suspect the same network behind at least five cases and more than 100 million ringgit of seizures this year, including a foiled attempt in June. Malaysia has controlled transshipment of U.S.-made advanced AI chips since 2025; this was at least its fifth interception this year.

My judgment: China is building a two-track model — IP and R&D stay onshore, capability and services are sold abroad. Export registration legalizes that path rather than closing it. In the short run, compliance capability will become a competitive advantage for Chinese models going overseas. In the medium run, the test is not token volume but whether overseas revenue can keep up. If by 2027 overseas revenue is still a rounding error next to usage, the "open source for market share" story will face repeated valuation challenges. The underrated variable is memory: compute can be multi-sourced and models can be registered, but HBM and advanced packaging remain concentrated in a few hands. That is China's hardest door.

Two more signals

In politics, Donald Trump said on Truth Social on October 27 that he would officially ban the use of "AI/Artificial Intelligence" in official documents and replace it with "Super Intelligence." It will not change the technology, but it will change the language of regulation. In capital markets, Anthropic filed for an IPO: $16.3 billion of first-half revenue, more than $80.7 billion of losses and a valuation that could reach $1.5 trillion, making it the first fully AI-focused public company. The filing also warns that AI could pose "catastrophic or existential risks" — Anthropic is both the biggest beneficiary of this financing boom and its biggest stress-test sample.

In closing

When compute can be multi-sourced, risk can be priced and models can be registered, the real dividing line is no longer whose model is strongest, but who can keep delivering inside someone else's rules.

Sources

• MyDrivers: OpenAI plans to buy Google TPUs
• CLS: OpenAI's Google TPU purchase challenges Nvidia
• Network World: Oracle adds Google TPUs to OCI
• CLS: Oracle expands Nvidia deal to gigawatts and adds Google TPU, Cerebras
• Yahoo Finance: TSMC to raise advanced packaging prices by up to 20%
• Yahoo Finance UK: AI chip boom tripled South Korea's growth
• Tom's Hardware: Nvidia is bankrolling the AI accelerator supply chain
• Maeil Business: China's AI boom drives race for Korean memory chips
• Eastmoney: Cambricon's latest results
• MyDrivers: Malaysia seizes 30 AI servers
• Chosunbiz: Broadcom leads $63B financing for Anthropic
• Maeil Business: BOK warns chip exports could fall $100bn
• Yahoo Finance: Chip selloff deepens on supply warning
• 199IT: HBM share of AI chip spending rises to 63%
• PR Newswire: 77% of enterprises have paid for ungoverned agents
• IBM: IBM Agent Identity is generally available
• SiliconANGLE: Red Hat introduces a sandbox for OpenClaw agents
• The Atlanta Journal-Constitution: AI bot accesses Georgia voting data
• TahawulTech: The kill switch problem
• Zaobao: DeepSeek and MiniMax complete overseas registration
• Reuters / WHBL: China's new export curbs draw a line around its AI crown jewels
• 21jingji: DeepSeek tops global weekly model usage
• Beijing Daily: Kimi K3.1 Pro at one-fifth the price
• CLS: China's State Council issues AI employment opinion
• Zhidx: iFlytek Spark X2.6 trained on domestic chips
• Yicai: Chinese models' overseas revenue lags usage
• The Independent: Trump says he will ban the phrase 'artificial intelligence'
• BNN Bloomberg: Anthropic files for potentially the biggest IPO ever
• BBC: Anthropic warns of AI's catastrophic risks
• Eastmoney: MiniMax open-sources M3.1 Flash