The short version: inference is not one market, but several battlefields.
Conclusion in one sentence: NVIDIA remains the standard, AMD is the most plausible second supplier, Qualcomm is the cheaper efficiency option - and Cerebras is the exciting specialist, but is currently only accessible to stock investors indirectly or as an IPO/private market topic.
Why inference should be valued differently from training hype
Training buys peak performance. Inference buys economics.
During training, whoever turns large models through the calculator faster counts. What also counts for inference: tokens per watt, latency, utilization, memory bandwidth, software integration and price per request. The boring part - cost curve - ultimately decides whether AI applications make money or just simulate nicely burning server rooms.
The software stack is the invisible chip.
NVIDIA benefits from CUDA, TensorRT-LLM, NVLink and system design. AMD counters with ROCm and open ecosystem. Qualcomm is trying to reduce developer friction through modular and heterogeneous platforms. Cerebras argues with single-device simplicity and extremely high on-chip communication.
Stock comparison: growth, profitability and price tag
| Company | Price | Market Cap | Revenue TTM | Revenue growth | P/E | P/S | FCF yield | Gross margin | Piotroski | Assessment |
|---|---|---|---|---|---|---|---|---|---|---|
| NVIDIA NVDA |
199.00 $ | 4,843 trillion. $ | $253.5 billion | 70.7% | 30.3 | 19.1 | 2.6% | 74.1% | 8/9 | Quality leader with a brutal moat - but not a bargain, more of the benchmark. |
| AMD AMD |
519.74 $ | $847.7 billion | $37.5 billion | 35.0% | 169.2 | 22.6 | 1.0% | 50.3% | 6/9 | Strong Challenger – but currently more “providing proof” than “collecting it cheaply”. |
| Qualcomm QCOM |
197.41 $ | $210.4 billion | $44.5 billion | 5.2% | 21.2 | 4.7 | 5.9% | 54.8% | 4/9 | Cheaper AI warrant with cash flow – less hype, more implementation risk. |
Four paths to AI inference
NVIDIA: AI Factory
GB200 NVL72 connects 72 Blackwell GPUs into one large NVLink domain. NVIDIA cites 30x faster LLM inference over H100 generation and 130 TB/s NVLink bandwidth in the rack.
Moat: software + networkingAMD: Open GPU Challenger
According to AMD, MI350 brings 288GB HBM3E and 8TB/s bandwidth. The investment case depends less on the data sheet and more on whether ROCm and hyperscaler deployments noticeably take market share from NVIDIA.
Moat: price / opennessQualcomm: efficiency thesis
Edge AI is Qualcomm's natural home advantage. The data center push with Dragonfly, Meta as an anchor customer and Modular as a software component makes the stock an AI option with real cash flow support.
Moat: watts per taskCerebras: Wafer Scale
WSE-3: 4 trillion transistors, 900,000 AI cores, 125 PFLOPS peak AI performance and 44GB on-chip SRAM. Technologically fascinating; as a stock but not directly tradable like NVDA/AMD/QCOM.
Moat: specialized architectureNVIDIA: solution, share, classification
Inference solution
Blackwell / GB200 NVL72, CUDA, TensorRT-LLM, NVLink, network + software stack
Investment thesis
The complete AI factory stack: chip, rack, network, software and developer ecosystem. For inference, NVIDIA is not just an accelerator provider, but the standard against which everyone has to explain why they are cheaper, more open, or faster.
Main risk
Valuation, supply chains/HBM, hyperscaler bargaining power and the risk that customers will build more of their own ASICs in the medium term.
AMD: solution, share, classification
Inference solution
Instinct MI350/MI355, ROCm, open ecosystem, EPYC server CPU integration
Investment thesis
AMD is the most credible secondary GPU supplier for hyperscalers that don't want to simply accept NVIDIA's pricing power as a law of nature. The MI350 series brings 288GB HBM3E and 8TB/s bandwidth; The lever lies in software maturity and real customer deployments.
Main risk
ROCm/developer momentum remains the bottleneck. High expectations are already priced in; the stock is unforgiving when margins and AI sales fall short of the narrative.
Qualcomm: solution, share, classification
Inference solution
Edge AI SoCs, Snapdragon/Hexagon NPU, Dragonfly C1000/AI300 data center push, modular software stack
Investment thesis
Qualcomm brings the efficiency gene from smartphones, automotive and edge devices to the data center. The meta-anchor customer and the modular topic are strategically exciting - just later in time and not yet as operationally proven as the smartphone license machine.
Main risk
Data center sales are delayed; Competition against NVIDIA/AMD plus its own customer ASICs. Smartphone cycles and China don’t suddenly disappear just because “AI” was stuck on the foil.
Cerebras: strong technology, but not a classic stock comparison
Cerebras is the architectural outlier in this report: wafer scale instead of many individual GPUs. The WSE-3 measures 46,225 mm², contains 4 trillion transistors and aims to solve communication within the chip rather than through many external connections. This is attractive for certain large models and latency profiles.
- Plus: radical architecture, high on-chip bandwidth, easy scaling story for special workloads.
- Minus: narrower ecosystem, manufacturing/yield complexity, customer diversification and currently limited exchange tradability.
- Investor assessment: interesting as a private/IPO-near pure play - but not with the same liquidity, transparency and balance sheet history as NVIDIA, AMD or Qualcomm.
Which stock fits which thesis?
Quality and moat investor
NVIDIA remains the cleanest quality bet. The price is high, but the growth, margins and ecosystem at least justify not only looking at the stock through the classic cyclical lens.
Challenger / market-share thesis
AMD offers more catch-up story, but also more evaluation stress. Interesting during setbacks or when new MI350 deployments really show volume and margin.
Cash flow + AI option
Qualcomm is less glamorous but cheaper. The stock needs patience because data center sales will only become visible later. There is a real basis for this.
Technology specialist
Cerebras is exciting for watch lists and IPO monitoring. Strongly differentiated as an inference solution, currently more difficult to model cleanly as a stock case.