Nvidia

Overview

Nvidia designs the accelerators, interconnects and rack-scale systems that carry the majority of the world's commercial AI inference traffic. The company does not manufacture its own silicon — it is a fabless designer that relies on TSMC for leading-edge process capacity, on SK hynix, Micron and Samsung for high-bandwidth memory, and on ODM partners for system assembly.

What distinguishes Nvidia in inference specifically is that it did not have to win inference as a separate market. CUDA, cuDNN, TensorRT and the broader library ecosystem were already the substrate that model training ran on, so trained models arrived at production deployment already targeting Nvidia hardware. Every purpose-built inference competitor described elsewhere on this site is, in practice, arguing that a narrower architecture can beat that default badly enough to justify porting cost and supply risk.

Nvidia's fiscal year runs to late January. Fiscal 2026, which closed on 25 January 2026, produced revenue of $215.9 billion, up 65% year over year, of which $197.3 billion came from the data centre segment. Fourth-quarter revenue was $68.1 billion with data centre revenue of $62.3 billion.

Chip and product families

Nvidia ships on a roughly annual architecture cadence, with mid-cycle refreshes that typically increase memory capacity and bandwidth rather than raw compute.

Role in inference

Nvidia hardware serves both stages of large language model inference, though not equally well. Prefill — processing the input prompt — is compute-dense and highly parallel, which suits a GPU almost perfectly. Decode, the sequential token-by-token generation stage, is usually memory-bandwidth-bound, and this is where specialist architectures make their case: a GPU spends a large fraction of its arithmetic capability idle while waiting on memory.

Nvidia's answer has been to attack the problem at the system level rather than concede the workload. Blackwell-generation rack systems allow prefill and decode to be disaggregated onto different pools of GPUs, so each stage runs on hardware configured for its bottleneck. Lower-precision formats reduce the bytes that must be moved per parameter. Larger and faster HBM raises the ceiling directly.

For most buyers the practical position is that Nvidia is the option that will run whatever model they choose, today, with known tooling. Specialist alternatives may be faster or cheaper on a specific model at a specific batch size, and that is a genuine and sometimes large advantage — but it is a narrower guarantee.

Strengths and key questions

Leadership

Jensen Huang co-founded Nvidia in 1993 and has been chief executive throughout, an unusually long tenure that has shaped the company's willingness to make long-horizon architectural bets. Colette Kress is chief financial officer. The company's research organisation under chief scientist Bill Dally has been a consistent source of the interconnect and numerics work that ends up in shipping products.

In December 2025 Nvidia agreed to license inference technology from Groq in a transaction reported at approximately $20 billion, its largest ever. As part of that arrangement Groq founder Jonathan Ross — previously a lead on Google's original TPU effort — and Groq president Sunny Madra joined Nvidia. The deal brought a non-GPU, deterministic inference architecture and its senior team inside the company.

Funding and financials

Nvidia is a public company and reports quarterly, so the figures here are audited or filed rather than announced.

The $500 billion figure is company guidance describing bookings visibility, not recognised revenue. It should be read as a demand signal that Nvidia has chosen to disclose, not as an audited result.

Customers and partners

Nvidia's disclosed customer base spans every major cloud provider, the large AI labs, sovereign compute programmes and a long tail of enterprise and neocloud operators. Blackwell-based systems are deployed across all major cloud platforms. Notably, several of the largest customers are also the companies furthest along in building competing internal silicon — Google with TPU, Amazon with Trainium and Inferentia, and Microsoft with Maia — which makes the relationship simultaneously commercial and competitive.

On the supply side, TSMC provides leading-edge wafers and CoWoS-class advanced packaging; SK hynix, Micron and Samsung supply HBM. Both are genuine constraints on how fast Nvidia can grow, and both are contested by every other accelerator vendor.

Technical positioning

Nvidia occupies the general-purpose end of the architecture spectrum. A Blackwell GPU will run a mixture-of-experts model, a dense transformer, a diffusion model, a recommendation model and a classical HPC workload, at good-to-excellent efficiency on all of them. An inference ASIC will beat it decisively on the workload it was designed for and may not run the others at all.

The strategic question is therefore about the stability of model architecture. If transformer-shaped decoding remains the dominant production workload for many years, the economic argument for specialisation strengthens over time. If architectures keep shifting — and the move toward reasoning models with long generated outputs and heavy KV cache pressure is a recent example of exactly that — flexibility retains real option value.

The Groq licensing deal is best read in this context. It gives Nvidia access to a deterministic, SRAM-heavy inference design that is architecturally unlike a GPU, at the point where the company's main competitive exposure is precisely the claim that GPUs are the wrong shape for decode.

That reading is no longer speculative. At GTC 2026 Nvidia unveiled the Groq 3 LPU, the first product of the licensing agreement, and positioned it inside the Vera Rubin platform as a dedicated decode co-processor: Rubin GPUs process long input contexts during prefill, and the SRAM-based LPU generates output tokens. It is a significant concession in principle — Nvidia is shipping a heterogeneous system in which the GPU is explicitly not the right device for half the workload. It is also a strong competitive position in practice, because Nvidia is now selling both halves rather than defending the GPU against a specialist selling one of them.

What to watch

• Rubin timing and specifications. Whether the annual cadence holds, and what the memory configuration looks like relative to Blackwell Ultra.
• Groq 3 LPU execution. Whether the decode co-processor ships to the Q3 2026 target, and whether disaggregated prefill/decode on Nvidia silicon delivers the cost-per-token improvement that specialists claim for their own designs.
• Samsung as a second foundry. The Groq 3 LPU is built by Samsung on 4nm rather than at TSMC, a notable diversification for a company whose supply is otherwise concentrated.
• Inference software releases. Dynamo and TensorRT-LLM improvements can move effective cost per token without any hardware change, and they are the cheapest lever Nvidia has.
• Hyperscaler mix shift. The share of hyperscaler inference running on internal silicon rather than purchased GPUs is the single most important long-term number for Nvidia's data centre business.
• Regulatory treatment of the Groq structure. Whether licence-plus-hire transactions attract sustained antitrust attention.
• Power availability. Increasingly the binding constraint on deployment is grid capacity and data-centre readiness rather than chip supply.

Sources

1. Nvidia — Financial Results for Fourth Quarter and Fiscal 2026. Revenue, data centre revenue and year-over-year growth figures.
2. Nvidia — Form 8-K, Q4 FY2026 CFO commentary (SEC EDGAR). Segment detail and supply commitments.
3. Nvidia — Form 10-K for fiscal year ended 25 January 2026 (SEC EDGAR). Audited annual figures and risk factors.
4. CNBC — Nvidia buying Groq's assets for about $20 billion (24 December 2025). Deal size, structure and leadership moves.
5. Groq — Non-exclusive inference technology licensing agreement with Nvidia. Company statement on the licensing structure.
6. ServeTheHome — Nvidia Q4 FY2026 earnings analysis. Independent reporting on segment performance.
7. Computer Weekly — Nvidia expands Vera Rubin platform, details Groq integration. Groq 3 LPU role within Vera Rubin.
8. Tom's Hardware — Nvidia's $20 billion Groq deal produces its first chip. Samsung 4nm manufacturing and prefill/decode split.

Financial figures are taken from Nvidia's own filings and earnings releases. Forward-looking demand figures are identified as company guidance. See the editorial methodology for how claims on this site are sourced and reviewed, and the contact page to submit a correction.

Related pages