The Inference Shift
Ben Thompson, Stratechery analyst, splits AI compute into two distinct kinds of inference. "Answer inference" has a human in the loop and speed matters. "Agentic inference" runs without a human, so latency is tolerable. The agentic side will be the bigger market - and it does NOT need Nvidia's speed premium.
Two kinds of inference. Two kinds of chip. Two kinds of datacenter. The market split everyone missed.
When a human is waiting, speed matters and Nvidia's HBM advantage is worth the premium. When an agent is doing overnight work, latency is fine. Slower DRAM, slower chips, slower locations all become viable. Thompson reads this as good news for Chinese fabs (no speed crown) and orbital compute (light-second delays acceptable). Bad news for Nvidia's pricing power on the agentic half.
For PMs: stop assuming one compute roadmap fits all your workloads. For execs procuring AI: split your inference RFPs into answer-tier and agentic-tier. For investors: the "Nvidia at any multiple" trade just got narrower.
⚡ Why this matters
- First clean framework for splitting inference compute into two markets with different economics.
- Reframes the SpaceX-Anthropic $1.25B/month compute deal and the Cerebras IPO as bets on the agentic-tier (latency tolerant) shift.
- Implies Nvidia's pricing power is asymmetric across workloads - a thesis Wall Street has not fully priced in.
🔍 What happened
- May 27, 2026. Ben Thompson publishes "The Inference Shift" on Stratechery.
- Three workload types: training, answer inference, agentic inference.
- Answer inference: human is the timer. Speed matters. HBM, low-latency interconnects, premium silicon.
- Agentic inference: no human waiting. Latency tolerable. Slower DRAM works. Cheaper chips work. Distance from the workload works.
- Geographic implication: agentic compute can live in remote, cheap-power locations - including orbital and Chinese fabs.
- Hardware implication: traditional DRAM, older process nodes, larger cluster designs all become competitive against premium Nvidia stacks for agentic workloads.
💬 Smart takes
- Ben Thompson (Stratechery): "agentic inference will be different than current inference and will change compute infrastructure because speed won't matter when humans aren't involved."
- Thompson on the market split: the answer-inference market is where Nvidia keeps its premium. The agentic-inference market is where it doesn't.
- Skeptic: Many "agentic" workflows still have a human waiting - chat sessions, code review loops, customer-facing agents. The clean two-bucket split may be over-tidy.
🧭 Where this goes
- Hyperscalers split their inference roadmaps into latency-tier and throughput-tier within 12 months.
- Specialty silicon vendors (Cerebras, Groq, Etched) reposition explicitly against the agentic tier in next investor decks.
- China's domestic chip stack (Huawei Ascend, Cambricon) finds its first real market in agentic inference where benchmark-leading speed doesn't decide the procurement.
- Nvidia carves out its enterprise pricing to defend the answer-tier while ceding ground on agentic batch jobs.
🎯 Implication
- For PMs: design your agent products to tolerate higher latency budgets. The compute layer will reward it.
- For execs: split your inference procurement into two RFPs. Pay the Nvidia premium only where a human is waiting.
- For investors: reprice every specialty silicon and neocloud bet against "latency-tolerant inference" specifically, not generic AI compute.