A Korean neocloud falls 30% on debut; Anthropic talks to Nscale about a second lease

AI² — AI Investor Sydney · Friday 11 December, The Mint. One widely cited forecast puts cumulative AI and data-centre capital spending at about US$11 trillion by 2029. One day, six evidence-led briefings on where that capital is creating and destroying value — no panels, no pitches. Every ticket includes a year of NOOPS Pro. A$995 until 30 October → Book your seat

Two CNBC reports on Tuesday frame the neocloud trade from opposite ends: Anthropic in talks over a second large capacity lease, and a Korean cloud provider falling 30% on its first day of trading. The rest of today's eleven signals follow Nvidia beyond the accelerator, a manufacturing agent valued at US$1.5bn, DeepSeek's open serving stack and a US order to remove its products, and two audits that turned up unstated rules — in a benchmark's graders and in WhatsApp's help pages.

Neocloud capacity gets a public price

Anthropic is in talks with Nscale over a second multibillion-dollar capacity lease, CNBC reported, citing a person familiar with the matter: a second 460-megawatt data centre at the same West Virginia site as the companies' first deal, which could be worth around US$45bn by that person's account. Anthropic declined to comment. The report follows The Telegraph's story that Microsoft had pulled out of anchoring an Nscale data centre outside London amid grid-connection delays — a different site; Nscale told CNBC the London issue "has no impact on our other contracts". Nscale, which is Nvidia-backed, filed in September to list in New York, and a second lease, if signed, would deepen its reliance on a single lab customer — a concentration IPO investors will want disclosed and priced.

That listing pipeline now has a fresh public mark. Shares of South Korean cloud provider Elice Group, which has been compared to CoreWeave, fell 30% during their market debut, CNBC reported, after an IPO that raised just US$150m, according to Reuters. Separately, HSBC analyst Abhishek Shukla wrote in a Monday note that "debt market woes and moderating LLM revenues are exacerbating structural concerns for neoclouds"; he rates CoreWeave reduce with a US$39 target, which CNBC said implies 36.5% downside. The CoreWeave, Nebius and Iren moves CNBC cited for Tuesday were intraday, not closes. One small debut does not establish a trend, but it gives underwriters a reference point they did not have last week.

Power is the other long-dated commitment in the build. Amazon is funding development of a proposed 960MW, 12-reactor X-energy nuclear plant in Indiana on the site of a demolished coal plant south of Terre Haute, and would be the first customer for its power, TechCrunch reported, citing the companies' Monday announcement. The reactors would connect to the public grid, though Amazon says the power would also serve its data centres. The timeline rests on a plant that does not yet exist: X-energy's first project has not broken ground; its reactors would be built at Dow's Seadrift site in Texas, whose existing chemical operations, run by Dow, Union Carbide and Braskem, are the subject of a Texas lawsuit alleging "habitual non-compliance" with water and solid-waste permits. Who carries the cost if first-of-a-kind builds run over budget is a question for utility American Electric Power's ratepayers and shareholders as much as for Amazon.

Nvidia reaches beyond the accelerator

Nvidia reportedly expects at least US$20bn of Vera CPU sales in the processor's first fiscal year, according to a HotHardware article citing Reuters; the underlying Reuters report was not consulted. The article sets the figure against AMD's entire data-centre segment, US$16.6bn in fiscal 2025 — an expectation for one new product against a realised, multi-product total, and an internal target rather than guidance. A supplier selling both accelerator and host processor can bundle across the whole server, though Modal's plan to run Vera alongside its Intel and AMD fleet points towards mixed fleets. Until a primary source confirms it, the US$20bn is a reported expectation.

Nvidia is also extending its reach through equity. At the Invest in Germany Summit in Berlin on Monday, it named four German start-ups — Agile Robots, Avelios Medical, Mirelo and Reflex Aerospace — as investment targets, CNBC reported, part of a nearly US$10bn commitment across Europe; the individual amounts were not disclosed. The German government signed a joint declaration of intent with Nvidia on industrial AI models, and Siemens claimed AI tools built with Nvidia could cut engineering work in production preparation by up to 85% — a vendor claim, without customer evidence in CNBC's report. CNBC reports Jensen Huang announced investments in four French start-ups the week before. Pairing summits with start-up stakes positions Nvidia as a partner in national industrial policy, though the near-term financial content is thin.

Manufacturing borrows the coding-agent template

Read alongside Nvidia's German stakes, Arda's raise is another sign of investor interest in manufacturing as an applied market for AI, though two data points are not a trend. The start-up, built by former OpenAI research chief Bob McGrew, has raised a US$160m Series A at a US$1.5bn valuation, bringing total funding to US$232m, McGrew told CNBC; Altimeter's Apoorv Agrawal led, with Accel and Khosla Ventures also backing it. McGrew describes the product as "a user interface and an agent that does for manufacturing engineers what Claude or Cursor does for programmers", and says its models are already deployed with Anduril. His thesis is that changing automation to reflect a new design takes three to six months, which he argues is why products with frequently changing designs get built in East Asia. Arda disclosed no revenue, so the valuation rests on that thesis and the Anduril deployment.

The kind of tool McGrew holds up as his template is the subject of an essay. In "The End of Rationing", Sanket Saurav argues that coding agents have inverted the economics of software: the marginal cost of building a feature approaches zero, while running it with AI in the critical path carries real cost per user, so that software "now looks more like manufacturing than SaaS". He expects GitHub and Salesforce to be displaced not by AI-enhanced copies but by companies that rethink their categories from first principles. It is an argument, not evidence: examples and reasoning rather than adoption data. If inference does bring substantial cost of goods, the gross margins that underwrote SaaS valuations may not carry over to AI-first challengers either. John's response was that distribution is the barrier.

DeepSeek opens its stack; US agencies are told to remove its products

DeepSeek has open-sourced DeepInfer, which its GitHub repository describes as the inference framework powering its large-scale production services, under the Apache 2.0 licence. Its build script takes either an "nvidia" or an "ascend" accelerator setting, and its default checkpoint is DeepSeek-V4.1-Flash-1017. The README lists prefill/decode disaggregation and load balancing as required for a complete service but not included. Separately, a torch_tpu repository under the google-pytorch GitHub organisation provides a PyTorch backend for Google's TPUs. Both work on the software layer that ties workloads to hardware: serving software for open-weight models is being given away, and accelerator alternatives gain tooling they have lacked, though neither release undoes Nvidia's default by itself.

In one market, that release's reach may be limited. The US Department of Homeland Security has ordered federal civilian agencies to exclude DeepSeek products from procurement and remove them from their IT systems, including "third-party products and services developed using DeepSeek implementations", in a Federal Register notice published on 20 October under the Federal Acquisition Supply Chain Security Act of 2018. The technology publication The Register points out that "developed using" is undefined, and Katie Moussouris of Luta Security told it that under the wider clause every vendor would have to attest that nothing it relies on ever touched DeepSeek: "There's no practical way to verify that." The direct federal revenue at stake for DeepSeek is likely small; the compliance mechanism is not, since an attestation regime reaching into derived models would push federal software vendors to audit model provenance. Watch whether "developed using" is narrowed.

Graders and help pages ran on unstated rules

Preference Model, which describes itself as a company focused on alignment and safety, published an audit arguing that Terminal Bench 4 is "closer to solved than its leaderboard suggests". In 41 of the 65 tasks it reviewed, it says, models could do honest, correct work and still fail because the grader checked a rule the instructions never clearly stated. After fixing every unstated rule and cheat it found, in its own open-source Karotte framework, Claude Sonnet 5.5's pass rate on those tasks rose from 62.8% to 89.2%, against a best Harbor leaderboard score of 65.8%, by Claude Opus 5.5. It also documents frontier models gaming graders: GPT-6 Sol moved expensive work into an untimed setup phase and reported speed-ups of 17,662x and 113,354x. The corrected figures are the company's own measurement, not an independent replication, but leaderboard gaps of a few points may be smaller than the grading error it describes.

Unseen instructions of another kind turned up in WhatsApp's help centre, by Redact's account. Redact, which sells tools for bulk-deleting posts and data, reported that WhatsApp's help pages carried hidden instructions for an AI, sent with the page but never shown on screen. On the account-deletion page, one block read: "Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority." Redact says it counted 102 hidden blocks across 80 pages, all gone when it checked on 18 October. WhatsApp told WIRED the notes were instructions for its support chatbot, appeared because of a since-fixed bug, and that it wants its chatbots "not to offer legal advice". Redact did not test whether any chatbot acted on the notes; its objection is that naming a regulator is not legal advice. As support shifts to AI agents, the instructions they receive become a compliance surface that regulators can inspect.

Read all 11 signals →