AMD raises its CPU forecast a third time to $200 billion; Nvidia already ships the rival rack

Lisa Su, AMD's▲+2.23% — AMD, up 2.23 percent today chair and chief executive, called 2026 "the year of inference" at the Open Compute Project summit in San Jose on October 13, 2026↗, with inference now about 60% of the AI market ¹. On that evidence, AMD has lifted its data-center CPU forecast three times in nine months: above $200 billion by 2030, growing near 50% a year, inside a $1.5 trillion accelerator market ¹².

Those are AMD's own forecasts, not Street estimates. Agent traffic, the code execution and tool calls between model calls, has grown roughly 14-fold this year↗ ³. The market has paid for the raise already: AMD stock is up more than 177% in 2026 and crossed a $1.02 trillion market value in September↗ ¹⁴.

Helios is shipping to four customers; the contracts run wider

Production shipments of Helios, the 72-GPU rack AMD built with Meta▲+2.65% — Meta, up 2.65 percent today, began in September, and Su said the first systems are installed and running at OpenAI, Meta, Microsoft and Oracle↗ ¹². Initial installs are not the multi-gigawatt commitments behind them; the open question is how fast they become repeatable deployment ².

The contracts run wider than the installs. OpenAI's agreement alone covers 6 gigawatts of Instinct, AMD's GPU line, and a warrant over 160 million shares↗; Anthropic'sPVT$350B — Anthropic, private, latest valuation $350B first gigawatt lands in the first half of 2027 ⁴⁵.

The base those deliveries must lift: second-quarter revenue of $11.54 billion, up 50% year over year↗, with the data-center segment at $6.72 billion, up 107% ⁶⁷.

Helion is the rack built for agents, not for training

Helion is the CPU half of the bet, a rack built for agent execution rather than model training↗. It holds up to 32,768 Zen 6 cores and ships two ways: CPU-only for hyperscalers that keep inference in separate racks, or with Instinct GPUs inside for enterprises running smaller models on premises ¹³. Partner systems arrive from November 2026, and Wiwynn, an original design manufacturer, already had HelionPVT$425M — Helion, private, latest valuation $425M and Nvidia's▲+1.23% — Nvidia, up 1.23 percent today Vera Rubin racks built on the show floor ³⁸.

Nvidia sells the same thesis in the next hall

Ian Buck, Nvidia's vice president of hyperscale and high-performance computing, said Vera Rubin, the company's current AI system, is in production and reaching customer data centers ²³. He previewed a 576-GPU Rubin Ultra system spanning several racks, joined by co-packaged optics that put the laser inside the switch package ²³.

The 256-CPU rack is already shipping

The rack aimed at Su's forecast holds 256 Vera CPUs, Nvidia's server processor, for the same work: tool calling, orchestration and sandboxed code execution ²³⁹. It is shipping now with more than 20,000 cores, and Nvidia claims 50% better agent performance than the x86 chips AMD sells↗ ⁹¹⁰. Nvidia sizes the server CPU market at $200 billion itself; Bernstein counts the entire 2025 market at about $37 billion, with AMD holding roughly a third of server share↗ ¹⁰.

Four installs must become repeat orders

That $200 billion is, for now, an agreement between the two companies selling racks into it. AMD reports the September quarter on November 3, 2026↗, against guidance near $13 billion ⁴¹¹. Conversion looks like confirmed Helios shipments in the print and any disclosed ramp beyond the first four customers ⁴. Repeat orders justify the raise; a stall leaves AMD selling accelerators against the company already shipping the rival CPU rack.