Good morning. The most important AI number in Monday’s news was not a valuation—it was a bill. Anthropic’s amended IPO filing puts its contracted compute and infrastructure commitments at roughly $895.4 billion, a reminder that frontier-model economics now look as much like heavy industry as software. That is the lead in today’s AI news today briefing, and the nuance matters: this is committed spend and lease exposure stretching into 2045, not cash already paid.

Elsewhere, Marvell agreed to buy Lumentum in a $115 billion all-stock deal aimed at the optical plumbing behind AI clusters; OpenAI offered a more flexible answer to data centers’ grid burden; and Abu Dhabi moved AI from pilot projects into government and energy operations. The common thread is unusually physical for a digital industry: power, fiber, contracts and organizational redesign now determine who can turn model capability into durable advantage.

AI news today: Anthropic’s $895.4 billion compute bill comes into focus

Anthropic’s November 2 amended S-1 discloses approximately $895.4 billion in purchase commitments and lease agreements, substantially all tied to compute and infrastructure. About $447.5 billion is expected before the end of 2030, most of the rest between 2031 and 2045, and $69.8 billion is due in the 12 months after September 30, 2026.

Those figures are contractual obligations—not conventional debt and not money already spent. Still, they reset the scale of the bet. Anthropic reported $33.24 billion in revenue for the first nine months of 2026, up 1,206% year over year, yet swung from $974 million of operating income in the second quarter to a $567 million operating loss in the third. The company attributed that reversal mainly to expensive short-term compute and non-cash donation-matching charges. For investors following Anthropic’s IPO path, the question is no longer whether demand is growing; it is whether financing, suppliers and margins can keep pace with obligations measured in decades.

Marvell and Lumentum put a $115 billion price on AI connectivity

Marvell and Lumentum signed a definitive all-stock agreement valuing Lumentum at an implied enterprise value of $115 billion. Lumentum shareholders would receive 4.5 Marvell shares for each share they own; after closing, current Marvell holders would own about 67% of the combined business.

The strategic logic is bandwidth. Lumentum brings lasers, optical transceivers and optical circuit switching; Marvell brings networking silicon, custom accelerators and interconnect. Together they want to sell more of the path between processors as copper reaches physical limits. Management projects a $550 billion addressable market and more than $1 billion in annual run-rate cost savings, but those are company forecasts. The deal still needs shareholder and regulatory approval and is not expected to close until the first half of 2028.

OpenAI proposes making giant compute campuses interruptible

OpenAI launched its Grid Protectors Program, starting with a planned Georgia campus. Its agreement with Georgia Power includes up to 1,000 megawatts of flexible demand response: the site could reduce electricity use when the grid is strained. OpenAI says it will pay the full infrastructure and service costs associated with its campuses.

This is a useful shift from arguing that AI load is economically beneficial in the abstract to specifying when a utility can curtail it. OpenAI also acknowledges that flexibility cannot replace new generation and transmission. The next test is operational: how often the campus can ramp down, for how long, and what service-level compromises result. As our recent look at AI capex and cash flow showed, infrastructure promises become meaningful only when contracts and measured performance catch up.

Abu Dhabi moves OpenAI into government and energy workflows

Abu Dhabi entities signed four agreements with OpenAI spanning government, energy, investment and adoption. The most concrete deployments include a machine-readable regulation pilot already in production for food-import rules and a partnership among ADNOC, G42 and OpenAI to apply AI to field-development planning and market forecasting.

The parties say officials retain final accountability in the government pilot and human oversight in the energy work. No financial terms were disclosed, so this is better read as an institutional deployment plan than a revenue event. What matters is the operating model: traceable rules, in-country inference and named business processes. If those controls survive real use, the agreements could become a template for sovereign AI adoption; if not, they remain polished memoranda with excellent lighting.

SUSE reportedly plans to cut roughly 10% of its workforce

The Register reports that SUSE told employees it would reduce its global workforce by approximately 10%. The report is based on an internal memo seen by the publication, not a public filing; SUSE is privately held. With a workforce of about 1,900, the plan implies roughly 190 roles, though the company did not publish that calculation.

CEO Dirk-Peter van Leeuwen reportedly pointed to automation and AI changing software-company workflows as part of the rationale. Leaders should resist the tidy conclusion that each removed job equals an AI replacement: strategy, costs and organizational redesign are tangled together here. What to watch is whether SUSE later provides regional details, restructuring costs and evidence that automation actually raises output rather than merely shrinking the denominator.

DeepSeek turns natural-language requests into Harness plugins

DeepSeek’s Harness team demonstrated Creator Mode on November 2, letting users describe a plugin in ordinary language and refine it through conversation. TechNode’s report points to examples including a paper reader, poster designer, dashboard, 3D workspace and video editor; some required several rounds of iteration.

The larger idea is more consequential than the demos: an agent becomes extensible by conversation, lowering the barrier between “use a tool” and “make a tool.” Harness is in public preview and open source, so builders can inspect the machinery rather than accept a sealed demo. The practical test will be whether generated plugins remain reliable, permission-aware and maintainable after the first impressive run.

Watch & Learn

Editor’s note: Cisco’s 10-minute “AI & ML Blueprint for Data Center Networking” explains why training and inference clusters need high-bandwidth, low-latency fabrics. It is a useful primer for product leaders and generalists who understand models but want a clearer mental picture of the network beneath them—especially before evaluating the Marvell-Lumentum logic.

AI, Translated: compute commitment

A compute commitment is a contract to buy or lease a specified amount of computing capacity, often over several years. Think of an AI company reserving a vast block of accelerator time and data-center infrastructure before it knows exactly how customers will use it. The reservation can secure supply and better long-term pricing, but it also creates obligations if demand, financing or technology changes. You should care because model economics increasingly depend on these contracts—not just benchmark scores or subscription revenue.

Try This Today: audit a filing with Grok

Goal: separate contractual facts from management forecasts in about 10 minutes. Grok’s current apps support PDF uploads and multi-file comparison, though limits vary by platform and plan, according to xAI’s official file guide.

1. Download a company filing or investor PDF from its primary source.
2. Upload it with the + control in Grok and identify the document and reporting period.
3. Ask for a three-column table, then verify every cited page against the original.

Copy-ready prompt: “Using only the attached filing, list material contractual obligations, historical results and forward-looking claims in separate columns. For every row, cite the exact page or section and quote no more than one short sentence. Flag estimates, conditions and non-GAAP measures. If the document does not support a claim, write ‘not established.’ End with five questions an investor should verify independently.”

One thing to remember

AI’s constraint is no longer only model intelligence. The decisive variables are increasingly contractual and physical: who has secured compute, power and connectivity—and who can still afford the promise when the invoice arrives.