What AMD's Ryzen AI Halo says about local inference economics

AMD shipped the Ryzen AI Halo in June 2026 at $3,999, undercutting NVIDIA's DGX Spark by $700 with a 14% edge on tokens per second. The price war matters less than what made it possible: local inference has turned into commodity hardware, and the number that counts now is tokens per second per dollar. Until recently, "local AI inference" meant a Mac Studio with 128GB of unified memory: one vendor, one platform, and a research curiosity for most builders. Now three vendors compete for the same workload: Apple's M5 Max, NVIDIA's DGX Spark and AMD's Ryzen AI Halo. All are under $5K, all have 128GB of unified memory, and all run frontier-scale models on a desk. More than 30 third-party mini PCs already ship with the Halo chip, starting at $2,399. The category went from research curiosity to commodity hardware in roughly nine months. Once that happens, performance matters less than economics. In the first half of 2026, AI moved from a FLOPS race to a race on tokens per second per dollar. Unit economics replaced raw performance as the number people compare. Current cost-of-ownership analyses put the break-even points roughly here. Below 20M tokens a month, managed cloud APIs win on total cost. Above 100M tokens a month, self-hosting on local hardware almost always wins. For continuous agentic workloads above 20% GPU utilization, local hardware pays for itself in under four months. One developer cut a $440 a month cloud bill to roughly zero by running a 235B-parameter model on a single Ryzen AI Max+ 395 mini PC. Six months earlier that workload needed a four-node Mac cluster. I see three consequences. 1. Agent economics flipped. Startups running production agent workflows defaulted to cloud APIs because hardware was for hyperscalers. Now $4K of capex can replace $400 a month of opex and pay for itself in ten months, so cloud by default is no longer the obvious call. 2. Cloud API providers face a new competitive surface. Below 20M tokens they win, above 100M they don't, and the middle of the market is contested. Providers will have to defend with something beyond raw inference: frontier model quality that local hardware can't match, enterprise compliance and audit infrastructure (SOC 2, HIPAA, data residency, SLA guarantees), and tooling that makes the API meaningfully easier than self-hosting. The ones that keep competing only on token price are in trouble within 24 months. 3. Enterprise infrastructure is converging on three tiers: public cloud for elastic and experimental work, private or local hardware for predictable high volume, and edge for low latency. Halo targets the middle tier, and that tier grows as agents move from prototype to production. So yes, Halo against Spark is a hardware race: faster, cheaper, more unified memory. There's a race at all because AI inference has graduated from a research metric to a finance line item. Every infrastructure team will have to decide in the next 18 months what to build and where to run it, and that decision will be made in the CFO's spreadsheet more than on a spec sheet.