🧭 Anthropic Commits $3.4B for 1.7 GW of Caterpillar Generators — Largest Known Power-Equipment Purchase by Any AI Lab
Reported by Bloomberg on November 8 (citing the Financial Times), Anthropic has agreed to pay $3.4 billion to take over generator purchase commitments from San Antonio-based OnePWR Solutions — covering 1.7 gigawatts of heavy-duty Caterpillar generators. This is the single largest known purchase of power-generation equipment by any AI laboratory, and the latest in a series of infrastructure deals that underline the scale of compute build-out Anthropic is funding ahead of its Nasdaq IPO.
How the deal came together
- Original plan: Anthropic initially negotiated to buy electricity from OnePWR, whose business model is to procure generation equipment and sell capacity to data-centre operators. Anthropic subsequently decided to absorb the generator orders directly — acquiring the equipment rather than the power output.
- Equipment: Heavy-duty Caterpillar generators, industrial grade; suited to data-centre primary or backup generation. 1.7 GW is roughly the continuous draw of a mid-sized US city.
- Context: This follows a cascade of infrastructure commitments: a $45B compute deal with Nscale (August 2026), a $1.3B private-credit data-centre loan in Texas, and the multi-GW Google/Broadcom TPU partnership announced in Q2.
- IPO use-of-proceeds alignment: The S-1 lists compute infrastructure as the primary use of IPO proceeds. Acquiring generator capacity — rather than leasing it — signals Anthropic plans to operate its own data-centre footprint at scale, reducing long-term lease dependency.
What this means for the AI power market
AI labs are now competing directly for generator capacity — the same equipment grid operators, utilities, and defence contractors need. By buying the generator orders outright (rather than contracting for power delivery), Anthropic locks in supply before Caterpillar's 24–36 month manufacturing queue fills up further. Expect hyperscalers and other frontier labs to announce similar vertical integration moves in Q1 2027 as the queue lengthens.
infrastructure
power
data centres
compute
OnePWR
Caterpillar
IPO S-1
🧭 Unified Claude Enterprise Experience — Chat and Cowork Merge into One Session, Default-On November 16
Anthropic has published a support article detailing the Unified Claude Experience for Enterprise, which eliminates the separate "Chat" and "Cowork" modes that Enterprise organisations have managed since the Cowork launch in Q1 2026. Under the new model, every conversation starts as a standard chat; when Claude determines that agentic work is required (file manipulation, code execution, shell commands), it automatically continues in an isolated cloud sandbox and returns completed outputs — with no manual mode switching by the user. The feature is available to opt in now via admin settings and becomes default-on for all Enterprise organisations on November 16, 2026.
Key design decisions worth understanding
- Isolated sandbox per conversation: When Claude begins agentic work, it runs in an encrypted cloud VM scoped to that conversation. Sandboxes persist for 30 days of inactivity, then are automatically deleted. Files are not shared across conversations unless explicitly exported.
- Cross-device continuity: Start a task on the desktop app, continue on mobile, finish in the browser — the sandbox state travels with the conversation. This is a meaningful change from Cowork, which previously had desktop-only agentic execution.
- Local access preserved: The desktop app still allows access to local folders, MCP servers, and local applications. Local execution is not removed; it is now an opt-in extension of the unified session rather than the default mode.
- Admin controls: Enterprise admins can enable early via Settings → Features → Unified Experience. Admins can also restrict which users or teams get agentic capabilities by toggling the sandbox permission tier.
- Companion November 5 release: The new Enterprise Security Center and inference hooks (beta) shipped alongside the early-access period — inference hooks let admins log or block specific API call patterns before they reach Claude.
Migration checklist for Enterprise admins before November 16
If your organisation uses Cowork-specific prompts or system-prompt instructions that reference "Cowork mode," review them before the cutover. Claude will now interpret all conversations as potentially capable of agentic execution — system prompts that reference mode switching may produce unexpected behaviour. The safe path: strip mode-conditional instructions and replace with capability descriptions ("you have access to file I/O and shell execution") that work regardless of mode.
Enterprise
Unified Experience
Cowork
agentic
sandbox
cross-device
admin
🧭 Ramp Data: AI Spending Stalls as Anthropic IPO Valuation Slides to $1.1–1.3T
A TechCrunch analysis of Ramp corporate spend data (published November 6) reveals a widening gap between AI usage metrics — which continue to rise sharply — and actual enterprise AI spending, which has plateaued. The divergence is explained by two forces: aggressive price cuts from AI providers, and increasingly sophisticated cost controls among the largest enterprise buyers. The data arrived mid-way through Anthropic's IPO roadshow, as Bloomberg reported that institutional investors were simultaneously revising the company's valuation from the ~$2T figure that circulated in early October down to $1.1–1.3 trillion.
The spending-usage gap — why it matters
- Usage up, spend flat: Ramp's lead economist Ara Kharazian notes that enterprise AI transaction counts rose quarter-over-quarter but average spend-per-seat declined. The delta is provider price cuts plus enterprises optimising to smaller, cheaper models for routine tasks while reserving frontier models for high-value work.
- Anthropic's headline numbers hold: H1 2026 revenue of $16.3B (up ~1,300% year-over-year); Q3 2026 revenue $16.8B; run-rate surpasses $30B — these figures make clear Anthropic is not affected by the flat-spend trend in absolute terms. The concern for roadshow investors is whether growth can sustain its pace once the initial enterprise adoption wave saturates.
- Valuation slide mechanics: Bloomberg (November 6) reported that IPO investors were weighing valuations as low as $1.1T — a 45% reduction from the ~$2T figure that circulated in early October — following a broad tech selloff and downward revisions to AI-infrastructure peers.
- Mizuho angle: Bloomberg reported on November 9 that Mizuho's underwriting role in the Anthropic IPO (raising up to $3.5B from investors including Japanese institutions) is positioning the bank to challenge Nomura's historic dominance of Japan-originated global IPO mandates. The Japanese retail tranche — added in the November 2 S-1/A — is a structurally interesting element: retail demand from Japanese investors may provide price support that offsets US institutional valuation compression.
What the spend-usage gap means for teams building on Claude
The enterprises in Ramp's data that held spend flat while growing usage are doing something worth emulating: routing tasks to the right model tier. If you are still running every task through Opus 4.7, you are likely overspending. Audit your agent calls by task type. Summarisation, classification, and short-form extraction almost always run acceptably on Haiku 4.5 at a fraction of Opus pricing. Reserve Opus for multi-step reasoning, long-context synthesis, and tasks where output quality directly affects revenue. The spend plateau is the market's proof that this tiered routing works.
IPO
valuation
enterprise spend
Ramp
model tiering
cost optimisation
Mizuho