AI Business Roundup: Anthropic, Taiwan Chips, Data Centers

Anthropic's finances, government rules on AI chips and AI spending inside large companies dominated this week's AI business roundup. Money moved in several directions at once: Anthropic reported large revenue and a proposed IPO-linked share pledge, Taiwan moved to criminalize unauthorized AI chip exports, and companies kept pushing AI budgets and data center capacity while the returns remain unclear.

Frontier AI Companies: Revenue, Shares and Safety Rules

Anthropic reported $16.8B in third-quarter revenue and an 85% adjusted gross margin, but those figures are unaudited, and the company's reported $2T valuation depends on verification. The company also plans to commit 5% of its shares to the Anthropic Public Benefit Fund, subject to shareholder approval and tied to its expected IPO. Separately, Anthropic has pitched a draft AI safety bill that would let US authorities block frontier models only above a threshold of 50 deaths or $1B in damage, as the unaudited Q3 numbers and the share pledge set out the company's financial position. The draft bill's high threshold means that blocking powers would apply only to the most severe harms.

In Massachusetts, a Senate bill would require independent evaluations of frontier models every 120 days, with developers paying the reviewers. The bill is still under negotiation. In the EU, EN 18286:2026 became the first harmonised standard under the AI Act and gives providers a presumption of conformity with Article 17.

Chips, Compute and Data Center Costs

Taiwan plans to make unauthorized AI chip exports to China a criminal offense. The move shifts compliance risk onto server makers and onto Southeast Asian transshipment routes, as described in our report on the plan. A reported $40B debt package for SpaceX, split between bank loans and bonds, would fund Nvidia chips for Memphis and orbital AI compute, though the terms are unconfirmed (SpaceX debt plan). Naver, meanwhile, plans a wholly-owned subsidiary to resell GPU capacity from its Nvidia-equipped Sejong AI data center, with a launch set for late 2026 (Naver's resale unit).

Meta's Q3 free cash flow turned negative at about -$2.12B while AI capital expenditure reached $34.9B, according to our Meta coverage. That spending is the clearest example of the build-out cost. The coverage of these projects does not address the costs for people living near the facilities, such as noise from on-site generation, emissions or cooling water use. Those costs deserve weight in any assessment of the Sejong, Memphis and orbital projects, and none of the summaries here quantify them.

Enterprise AI: Spending Versus Returns

Accenture now centrally funds $29 per employee per month for Claude, with usage above that level charged to project budgets (Accenture's Claude spending cap). Gartner's survey found that only 13% of CIOs report significant AI value, while 58% face cost pressure (the Gartner value gap). SAP's cloud backlog reached €24.1 billion in Q3 2026, up 26%, with Joule AI cited as a driver, although the share attributable to AI is not quantified (SAP's backlog growth). AI is cited in about 21% of 2026 US job cuts, yet a Federal Reserve Bank of Atlanta paper finds most firms see no AI productivity effect (the job cut attribution story). Proofpoint's acquisition of Varonis values the target at $6.3B, or 8.4x revenue and 53x free cash flow (the Proofpoint Varonis deal).

Publishers, Models and Robotics

Google added CNN and Reach to its AI licensing deals, while USA Today Co sued OpenAI for $250M (Google's licensing expansion). DeepSeek upgraded V4.1 Flash to DeepSeek-V4.1-Flash-1017 with unchanged API pricing and reduced runtime (the DeepSeek upgrade). Google also released TorchTPU as open source, which lets PyTorch apps run on TPUs without rewrites, though the benchmarks are Google's own and remain untested (the TorchTPU release). Boston Dynamics says its Atlas humanoid now takes a day to build and targets one hour by 2027, ahead of a Hyundai factory rollout in 2028 (the Atlas build-time plan). In London, the robotaxi rollout has stalled while TfL guidance is pending, and the London Assembly seeks fleet caps and TfL oversight (the London robotaxi story). KAI's satellite AI pilot on the BEE-1012 CubeSat tested one onboard fault-response loop, with cost savings and full autonomy still unverified (the KAI pilot).

Watch whether the Anthropic financials are audited before the proposed IPO and share pledge move forward.

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.