AI ASICs — Strategy
Updated 6/19/2026
Where AI ASICs is heading over the next 12 months, grounded in product-axis evidence and verbatim demand from the last 90 days. The judgment column is the engine's read — operators verify and refine.
Product trajectories
Groq LPU inference appliance — AI inference accelerator ↗ rising
Opportunity: Sovereign-AI demand outside US hyperscalers wants dedicated low-latency inference fabric.
Groq is the clearest pure-play inference-ASIC bet with a real sovereign deal anchor; hiring tilts into compilers and capital-markets (financing the buildout), not raw RTL — they have the chip, now they are scaling commercial+capex.
AWS Trainium (Trainium / Trainium2 custom AI silicon) — AI accelerator (hyperscaler ASIC) ↗ rising
Opportunity: Hyperscaler displacement of Nvidia GPUs with bespoke ASICs; Anthropic being pushed onto Trainium creates a guaranteed inference workload.
Trainium is the most-cited custom hyperscaler ASIC across demand chatter; Marvell is the locked-in design-services partner. Strong wind-direction product but most A1/A4 evidence sits inside AWS/Annapurna (not in the listed companies).
Google TPU (custom hyperscaler ASIC) — AI accelerator (hyperscaler ASIC) ↗ rising
Opportunity: Investor narrative shift toward profitable inference on custom ASICs vs renting Nvidia GPUs.
TPU is the canonical 'ASIC beats GPU' reference point in retail and investor discourse; Broadcom is the disclosed silicon partner monetizing it. Confirmed product-market fit; the public chatter is itself the trend signal.
AMD Instinct MI450 GPU — AI accelerator → steady
Opportunity: Hyperscaler need for a credible second AI accelerator vs Nvidia (fact f953dea3-974f-4b3e-9c78-bdb57cb6a88f: 6 GW OpenAI commitment plus Oracle 50K MI450 line).
MI450 is the single most-anchored ASIC product in the evidence: two named hyperscale customers (OpenAI 6 GW, Oracle 50K) plus a $5B revolver and an explicit roadmap (MI300→MI325→MI350→MI430X→MI450). AMD is the strongest non-Nvidia bet in evidence.
Cerebras Wafer-Scale / Condor Galaxy inference cluster — wafer-scale AI accelerator + inference cloud → steady
Opportunity: Demand for >2,500 tok/s inference on frontier open models (fact 8bc85f5e: Llama 4 Maverick).
Cerebras is the only fully-stacked ASIC company in evidence with all five buckets covered: capex, anchor customer, US DC expansion, named inference products, and hiring an ASIC Architect to keep iterating. Strong hot signal.
Compiler / MLIR software stack for AI ASICs — ML compiler & runtime → steady
Opportunity: Every ASIC needs a usable software story; fact b6ac96cb notes MLIR-based ML compiler skills are a recurring requirement.
The wedge that decides which AI-ASIC startup survives. Compiler hires are concentrated at the inference-pure-plays (Groq, Tenstorrent) — the moat is portability of HF/PyTorch workloads onto custom silicon, not silicon itself.
What the market is asking (last 90d)
- Will AI replace ASIC Verification Engineers?
- ASIC vs GPU : Which is better for AI work load?
- Paper Tape Is All You Need – Training a Transformer on a 1976 Minicomputer
- Ask HN: What is your (AI) dev tech stack / workflow?
- Will ASICs lead in AI inferencing?
- Resistance training load does not determine hypertrophy
See the Products and Hiring modules for the full landscape and who's investing in which direction.