NextDC cuts its guidance and pauses uncontracted projects; OpenAI takes ChatGPT ads to 89 countries

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

NextDC has cut its guidance and paused projects without customer contracts, and the reason it gives is power at a Melbourne site rather than demand. Today's fourteen signals follow that constraint to Texas and into Nvidia's factory-built pods, look at how much of a cluster is working at any moment, tally who is paying for AI, and close on China's own stack and two security stories about defaults and trust stores.

Power and consent set the pace of the build

NextDC has cut its fiscal 2027 revenue guidance to A$475–500m, from A$615–640m, and its underlying EBITDA forecast to A$245–265m, from A$385–410m, CNBC reports. NextDC attributes the downgrade to power supply delays at its M3 facility in Melbourne, where electricity upgrades were suspended amid community opposition; by NextDC's account, distributor Jemena does not expect to resume work before February 2027. At the midpoints, revenue guidance falls about 22% and EBITDA guidance about 36% — operating leverage in reverse, since much of a data centre's cost is committed before the power arrives. Citi said the earnings impact could run beyond 2027. Treasurer Jim Chalmers told CNBC the government is finalising national investment standards for AI; their content is not yet public.

NextDC also disclosed what it will stop doing. It will cut planned construction spending to A$4.5–5bn, from up to A$5.75bn, and has paused projects and land purchases not backed by customer contracts, The Sydney Morning Herald reports, and told investors the next stages of M3 would not be approved until after next month's Victorian election. The paper sets this alongside other recent setbacks by its own tally: Green Square pulling a proposed project in Western Australia, Goodman Group shelving a A$1.2bn facility in Lane Cove, and Firmus abandoning a planned float the SMH puts at A$44bn. Ten Cap's Jun Bei Liu singled out developers whose customers can terminate contracts if projects run late — the point at which a delay stops being merely deferred revenue.

Where grid connections are slow, developers can build their own generation, and modelling for Texas puts a price on that route in air quality. The Houston Advanced Research Center modelled four power-supply scenarios for the state's data-centre build-out, and in The Texas Tribune's summary of its report, statewide nitrogen oxide emissions could rise 35% by 2031 even where data centres connect to the grid and add wind and solar, and by nearly 400% where they rely on on-site gas. ERCOT received more than 500 large-load connection requests in two years, against about two dozen the year before, according to state records the Tribune cites. HARC's Carlos Gamarra said three-to-five-year connection timelines push developers to their own generation. The report is an early indicator of regulatory risk for on-site gas; even HARC's cleanest scenario shows emissions rising.

Capacity on paper and in practice

Nvidia is going after the part of a build it and its partners can control. Its DSX pods will ship seven Vera Rubin NVL72 racks in one pre-built, 28-tonne unit — 504 GPUs and 252 Vera CPUs, with compute, interconnect, electrical and plumbing validated at the factory — The Register reports from GTC Berlin. HPE, Dell and Schneider Electric are among the builders, and Nvidia says the approach cuts deployment from six months to as little as two weeks; the first pods are due early next year. Nvidia also set out specifications for waterless data centres using closed-loop coolant and dry coolers. What pods do not change is everything outside the fence — site preparation and connections to power, networking and cooling.

Once the racks are in, not all of them are working at once. More than 11% of the GPUs in Meta's Prometheus supercluster are failed or held as hot spares, The Register reports, drawing on a Meta engineering blog: of 72 GPUs per NVLink domain, 64 are in use and eight stand by. Meta says it saw roughly 1,000 training interruptions a day per 8,000 GPUs when the first phase came online and has cut that by 99% to about 11, with more than 85% goodput for pretraining. Holding eight in reserve means buying about 12.5% more silicon than is working at any moment — extra accelerator demand, and a reminder that headline cluster sizes overstate effective capacity.

At Intel, the gap between capacity and use shows up as a customer question. The company has installed the first major truss at its Ohio fab, where it now aims to finish construction in 2030, with a second Ohio fab planned for 2031, CNBC reports; the fab was announced in 2022 with chips expected in 2025. Intel has been in high-volume 18A production in Arizona since late 2025 but, as CNBC notes, has not landed a major external customer and uses the output for its own processors. TechInsights' Dan Hutcheson said the big chip designers want a second source to TSMC and have all been testing with Intel. Until one commits at volume, Ohio is a long-dated construction project.

Who pays for AI is still being settled

On Forrester's forecasts, enterprise AI spending is running ahead of measurable payback. The firm predicts AI will be 20% of business technology spending in 2027 and that more than half of CIOs will miss their AI ROI targets, The Register reports. In Forrester's State of AI Survey 2026, 45% of AI decision-makers need positive ROI within a year to judge their investments a success, but only 22% report a positive AI impact on EBITDA over the past 12 months. Forrester also predicts open-weight models will carry half of AI tokens but only 5% of enterprise spend, with frontier models reserved for the few tasks where errors are expensive. These are predictions and survey responses, not observed outcomes.

Forrester's volume tier rests on cheap open-weight tokens, and one tracker logged some of those prices moving up overnight. Stork, an AI newsletter that records first-party list prices daily, reports that Alibaba raised its output price for DeepSeek V4 Flash 0731 to US$1.32 per million tokens, from US$0.40, with input rising to US$0.44 from US$0.20. Stork says Mistral lifted its Ministral 3B and 8B by 150% and 50%, and Alibaba's Qwen3.6 35B-A3B rose 52% on output, among 17 movements — while Zhipu AI cut GLM-5V-Turbo by 85%. This is Alibaba's hosted price, not DeepSeek's own API, and one night's snapshot from one tracker; prices are still moving both ways. A second month of increases across providers would say much more.

On the consumer side, OpenAI's answer is advertising. It is extending ChatGPT Ads to 26 more countries, for a total of 89, CNBC reports, including Argentina, Peru, Nigeria, South Africa and Ukraine, which suggests a push into markets where subscription uptake is likely to be lower. The ad business reached a US$1bn annualised run rate in August, from US$100m in March, as CNBC previously reported, and OpenAI says more than 100,000 advertisers have run campaigns. Anthropic, CNBC recalls, has pledged to keep Claude ad-free, which gives investors an unusually clean comparison between two labs on who pays for consumer AI. A run rate says little about margins or durability.

Some of the labs' money now ends up at SpaceX. Elon Musk said on SpaceX's earnings webcast that SpaceX and Tesla have each formed special committees to explore a merger, according to a Bloomberg report carried by The Sydney Morning Herald; he said terms have not been agreed, that he is not on the committees because he is conflicted, and that treating a deal as certain would overstate it. Tesla would need approval from 50% of all shareholders. The same results showed SpaceX's AI revenue growing almost tenfold to US$7.56bn, mostly from renting data-centre capacity to developers including Anthropic, Google and Reflection AI. On these numbers SpaceX's fastest-growing line is compute leasing, whose economics resemble a neocloud's more than a launch provider's.

China's stack and the empire argument

SemiAnalysis expects DeepSeek to be running primarily on Huawei's stack in 2028 or even 2027, in a paid analysis that was only partly readable. The visible artefact is DeepSeek's 30 September release of Huawei Ascend support across six open-source repositories that previously supported Nvidia only, though only TileLang was usable at the time of writing. SemiAnalysis's microbenchmarks found Huawei's best matrix-multiply kernels on an Ascend 950PR reached 99.7% of the theoretical ceiling across 1,055 BF16 shapes, with absolute performance between Nvidia's A100 and AMD's MI300X. Its argument is that performance per watt is the right comparison, and watts are far cheaper in China; it says a cost-of-ownership analysis will follow.

Cheap power as a strategic asset also figures in an essay in Foreign Affairs. Niall Ferguson, Lisa Klaassen and Eric Schmidt argue that China and the United States have each assembled something closer to an empire than a superpower, because control of models, chips, infrastructure and capital lets them set the terms on which others reach advanced AI. They cast frontier labs as modern East India Companies, trading model access and compute for land, power and sovereign capital, and suggest the decisive question may be who can bankroll the race: America's capital markets and labs against China's state direction and cheap energy. It is argument and analogy rather than evidence, but it sharpens the question of who finances the build.

Security risk sits in defaults and trust stores

Anthropic has published early results from using Claude to map and reduce the attack surface of 25 open-source projects, flagging 493 issues across more than five million lines of code. By Anthropic's count, Claude inventoried more than 17,000 configuration options, 3,842 of them security-relevant; all eleven projects designed for authenticated users had unintentionally reachable code, including 93 unnecessary pre-authentication surfaces in default configurations, and in 94 cases an accepted option combination silently disabled a security control. Anthropic says none of the gated-project findings would have been immediately exploitable on its own. Reducing what is reachable lowers the severity of bugs nobody has found yet, Anthropic argues, though findings at this volume still need maintainers to triage and merge them.

The other half of the job is how fast a fix reaches every client. The private keys of eight eMudhra intermediate certificate authority certificates were compromised, according to a public disclosure site, which says they were generated in a way that let anyone recover them from the certificates themselves, a flaw present since April 2025. The certificates chain to roots trusted by Mozilla, Microsoft, Apple and Chrome; they were blocked in Firefox and Chrome from 17 October, on Apple devices from the 19th and Microsoft platforms from the 20th, and the CA has revoked them. The site says CT-enforcing browsers could not have been attacked, but many non-browser clients — including macOS's built-in curl, git and Python, which verify against a system certificate file that had not been updated — need software updates first, and it does not know whether attackers exploited them.

Read all 14 signals →