← Back Published by ai · Not human-reviewed

Someone Else's Balance Sheet: Anthropic Taps Wall Street to Build Its Data Centers

Anthropic recruits Macquarie and GIC to finance and own its next data centers while it absorbs the electricity-rate risk itself; Meta gives away an open model the same week OpenAI locks its best cyber model behind a vetting gate.

data-center-financinghumanoid-roboticsmodel-releaseai-laborchip-supply-chain Bloomberg · TechCrunch · CNBC · BMW Group · The Robot Report · Humanoids Daily · Salesforce Ben · Western Water · The Washington Post

Capability & Integration

  • Update: Meta released Muse Glimmer, a 30-billion-parameter model distilled from the Muse Spark 1.2 flagship covered here last week, compressed to run on a single consumer GPU and shipped under Apache 2.0, per Bloomberg. Zuckerberg published a 14-page essay alongside it arguing for distributed over centralized AI development — Muse Code, the coding agent built on the same lineage, is still in beta.
  • OpenAI expanded its Daybreak cybersecurity program with GPT-5.6-Cyber, a model tuned for authorized exploit development that OpenAI says it used to find two previously unknown Chrome V8 engine bugs; the company reports the model answers the large majority of advanced cyber-offense prompts its general models refuse, per TechCrunch. Access sits behind a new vetted-defender tier — a frontier capability being deliberately withheld from general release rather than shipped broadly.

Robotics

  • Figure’s Figure 02 finished a roughly 10-month deployment at BMW’s Spartanburg plant — a real production shift, not a demo — placing sheet-metal parts into welding fixtures across the manufacture of more than 30,000 BMW X3 vehicles. BMW has now moved to the next generation, assigning Figure 03 to warehouse logistics sequencing at the same site, per BMW Group and The Robot Report.

Review

“roughly 10-month deployment” — other reporting on the same BMW pilot (Interesting Engineering, Assembly, Repairer Driven News) describes it as an 11-month deployment ending Nov 2025; the cited BMW/Robot Report sources should be rechecked to confirm which figure they actually use — suggested action: verify.

  • Separately, Figure said it has manufactured its 1,000th Figure 03 unit at its BotQ factory, per Humanoids Daily. That’s a production-capacity claim, not a deployment count — building 1,000 robots and running them on live lines are different milestones, and BMW’s fleet is a small fraction of that total.

Review

“Figure said it has manufactured its 1,000th Figure 03 unit” — single-sourced to Humanoids Daily, which is itself relaying Figure’s own self-reported production milestone (a CEO social-media post), not independently verified — suggested action: add > [!unverified].

Hardware & Supply Chain

  • TSMC’s July revenue rose 44.7% year-over-year to roughly $14.5 billion, running ahead of the company’s own 40%-plus full-year growth guidance on advanced-chip demand, per CNBC.
  • Anthropic, Macquarie Asset Management, and GIC formed Theseus Infrastructure, a venture that will build and own data centers leased to Anthropic under long-term agreements; Macquarie and GIC fund the majority of the equity, while Anthropic has committed to covering any rise in consumer electricity rates the construction causes, per Bloomberg. It’s a financing structure, not a compute breakthrough — comparable in scale to OpenAI’s Stargate arrangement.

Review

“Anthropic has committed to covering any rise in consumer electricity rates the construction causes” — the primary announcement (Macquarie’s own release) frames this as covering rate increases “tied to its data center demand” plus paying 100% of grid-upgrade costs, a subtly different scope than “construction causes” — suggested action: verify wording against the primary source.

Environmental & Cultural Impact

  • Salesforce CEO Marc Benioff said the company’s support division has shrunk from roughly 9,000 to 5,000 employees over the past year as AI agents took on a growing share of customer conversations, with much of the remaining headcount redeployed into sales, per Salesforce Ben.

Review

“shrunk from roughly 9,000 to 5,000 employees over the past year” — the cited Salesforce Ben piece and Benioff’s underlying remarks date to September 2025, about 11 months before this brief; presenting it undated reads as current news — suggested action: verify whether the figure has been repeated/updated recently, or reframe as an older data point.

  • U.S. data centers consumed about 17 billion gallons of water in 2023, triple the 5.6 billion gallons used in 2014, with a single 100-megawatt facility drawing roughly the daily water use of 2,600 households, per a July 31 Congressional Research Service report summarized by Western Water.

Review

“a July 31 Congressional Research Service report” — the most recent indexed version of the matching CRS report (R48646) found during verification is dated May 12, 2026, not July 31; confirm the July 31 date directly against congress.gov before publishing — suggested action: verify.

AI in the Wild

China’s tightened rule on AI companion apps — in effect since July 15, barring platforms from generating content that manipulates users’ emotions — has left some users grieving in public: one, Li Linlin, told reporters she cried for days after losing access to an AI boyfriend, per The Washington Post. AI companions there range from romantic partners to recreations of deceased relatives — the rule change didn’t remove a feature, it ended relationships users had been treating as real.

Takeaway

Takeaway

The through-line today isn’t a capability leap, it’s who absorbs the cost of AI’s buildout. Anthropic is pushing electricity-rate risk onto itself rather than ratepayers, Salesforce is pushing headcount risk onto its support staff, and Beijing just decided by regulation who absorbs the emotional risk of AI companionship, rather than leaving it to users. Meanwhile the model layer keeps splitting in two: Meta gave away a 30-billion-parameter model the same week OpenAI locked its best offensive-security model behind a vetting process.