Wednesday evening produced an unusual compression of news from Nvidia: a fiscal quarter that crushed already-elevated forecasts, an expanded AWS hardware commitment running deep into 2029, and a separate Reuters-confirmed report that Nvidia has agreed to acquire Hugging Face for $12.9 billion. Read together, these moves signal that Nvidia is now attempting to own the full chain from chip design through model distribution. The competitive and regulatory consequences of that attempt will take months to settle.
Nvidia Q2 FY2027: $96 billion, $108 billion Q3 guidance, 70 per cent FY2028 growth
Nvidia posted $96 billion in revenue for its fiscal second quarter, more than doubling the year-earlier figure and arriving nearly $4 billion above the $92 billion analyst consensus. Data centre revenue reached $89 billion, with $48.7 billion from hyperscalers and $40.3 billion from AI clouds, industrial, and enterprise deployments. Gross margin held at 75 per cent, unchanged quarter-on-quarter, though management flagged rising memory and wafer costs as a continuing headwind.
Q3 revenue guidance came in at $108 billion, plus or minus 2 per cent. The more consequential signal was Jensen Huang's preliminary FY2028 colour: approximately 70 per cent revenue growth, described explicitly as supply-constrained. Demand, Huang told analysts, is materially higher than what Nvidia can currently ship. His shorthand from the earnings call, "Now, compute is revenue," is a useful compression of where the AI capital cycle stands at present.
Alongside the earnings release, AWS and Nvidia published a joint announcement covering an additional 2 million GPU deployments, running from this quarter through fiscal Q2 2029. The agreement also covers Nvidia's new Vera CPU at scale. Vera is Nvidia's first ground-up central processor, competing directly with Intel and AMD. Some Vera deployments will be integrated with the forthcoming Rubin AI chip; others will run standalone. Management guided Vera CPU revenue to more than double in FY2028, placing Nvidia into a market both Intel and AMD had held without meaningful competition for decades.
Nvidia agrees to acquire Hugging Face for $12.9 billion
The Information reported on 26 August, confirmed by Reuters the same day, that Nvidia has agreed in principle to acquire Hugging Face for $12.9 billion. The deal has not yet closed and may still fall through, pending final terms and regulatory review. Hugging Face operates the largest public repository of open-weight models, datasets, and inference tooling, and hosts the model cards and weights for most of what the independent AI development community runs.
The strategic logic is direct: every open-weight model hosted on Hugging Face runs overwhelmingly on Nvidia hardware, and owning the distribution layer gives Nvidia visibility into what is being built before it ships, the ability to optimise platform tooling for its GPU stack, and a commercial relationship with the developer community it currently reaches only at a distance. The risks are equally clear. Antitrust scrutiny in both the US and EU is a near-certainty given Nvidia's existing dominant position in AI silicon. Rival chip suppliers, AMD and Intel among them, have a material interest in a neutral distribution platform that would no longer exist if the deal closes. The same applies to Anthropic and OpenAI, both of which are building custom inference silicon to reduce Nvidia dependence and would, post-acquisition, be routing their open-source-adjacent activity through a Nvidia-owned hub.
IBM Granite 4.2: open-weight agentic reasoning for enterprise, Apache 2.0
IBM released Granite 4.2 on 25 August, a family of open-weight decoder models in three sizes (3B, 8B, and 30B parameters), all under the Apache 2.0 licence. Each model supports a 512,000-token context window and a switchable thinking mode, letting operators trade latency for reasoning depth at inference time without changing model versions. The 8B and 30B variants were additionally trained through agentic reinforcement learning, meaning they were placed inside live software-engineering, terminal, and web-search environments and graded on task outcomes rather than next-token prediction.
All three models use the OpenAI function-calling format natively, so they integrate into existing agentic harnesses without adapters or shim layers. Weights are available on Hugging Face, GitHub, and Ollama; quantised GGUF variants down to Q4_K_M are published for local serving. For an enterprise team that needs a reasoning agent it can run on-premise, audit end-to-end, and clear under a single reading of the Apache 2.0 terms, Granite 4.2 now sets the benchmark against which proprietary alternatives need to justify their cost.
OX Alpha: an anonymous stealth model at 80 per cent on DeepSWE, free window closes today
An unnamed model called OX Alpha appeared on OpenRouter on 20 August under the designation "stealth model," with a 1 million-token context window, coding and long-horizon software tasks as its declared purpose, and a one-week free preview. That window closes today. Bloomberg first covered the model on 23 August; independent researchers subsequently benchmarked it against the DeepSWE coding evaluation suite and found it completing 8 of 10 tasks, an 80 per cent pass rate, compared with Claude Fable 5 at 65 per cent and GPT-5.6 Sol at 52 per cent.
The identity of the builder remains officially unconfirmed. Developers applied tokenizer fingerprinting and found a consistent 75-token offset from Zhipu AI's GLM-5.3, placing high confidence on OX Alpha being a pre-release GLM-5.x variant. The approach, deploying an anonymous model on Western developer infrastructure during a limited free window, yields real-world performance data under production load before any formal announcement. If the identification is correct, Zhipu AI is operating at parity with or ahead of the named closed frontier on code at the time of publication.
The thread connecting these four items is the question of who controls the AI stack, and that question is no longer theoretical. Nvidia's earnings confirm demand is running well above what any competitor can match in the near term; the Hugging Face move would extend that structural position into the software and distribution layer. Against that backdrop, IBM's Apache-licensed release and the Chinese stealth-model pattern both represent meaningful counterpressures. Operators deciding where to build this year are, in effect, placing a bet on how that consolidation resolves.