Intel’s 28.58% Fall Meets Google’s Reported TPU Order

Intel shares closed down 28.58% after weak guidance, their worst one-day fall since the 1970s. Yet Google had reportedly ordered Intel to manufacture more than three million of its TPUs for 2028. Investors were punishing the company that sells the processor; a prospective customer was seeking the company that could build someone else’s.

Intel’s share-price collapse did not settle the value of its factories, just as a reported order for future production does not prove those factories can deliver it.

AI is turning compute from a CPU-centered product market into a coordinated capacity market. Intel’s long advantage combined stewardship of x86 with control of its manufacturing roadmap; buyers of AI infrastructure now assemble accelerators, host CPUs, memory, interconnects, packaging, software, and fab capacity across companies. Intel can remain central to that work, but as a competitive supplier to systems whose designs it does not control—not by restoring sole command of the platform.

x86 became an interface Intel has to share

The IBM PC’s adoption of x86 helped create an installed base large enough to reward decades of compatible processor development. Enterprises could keep software running across successive machines, and Intel could improve chips while relying on that continuity. Compatibility was both a customer benefit and a boundary around the market.

Intel still treated the boundary as defensible in 2017, when it warned that x86 emulation on Windows laptops using Qualcomm processors could present patent risks. By 2024, Intel and AMD had formed an x86 ecosystem advisory group to maintain a unified instruction set amid competition from Arm and RISC-V. Intel had not abandoned x86. It had joined its principal rival to preserve the common language both needed customers to keep using.

That distinction matters when reading the latest revenue forecasts. AMD projects roughly $17 billion in fourth-quarter revenue against Intel’s $15 billion–$16 billion outlook; the comparison raised the prospect of AMD becoming the largest x86 vendor by revenue for the first time. Both figures are forecasts, and companywide revenue does not establish x86 processor revenue or market share. The durable fact is narrower: an Intel customer can retain x86 compatibility without buying its CPU from Intel.

The server buyer now chooses the assembly, not just the processor

A CPU still runs essential work around an AI accelerator. That role has attracted more competitors, not fewer: Nvidia, Qualcomm, MediaTek, Arm, Google, Amazon, Intel, and AMD are all part of the widening contest for AI-server CPUs. The buyer’s question is no longer simply which processor replaces the last one. It is which combination can run a workload and arrive in usable quantities.

In 2024, Intel joined Google, Microsoft, Meta, and others in forming the Ultra Accelerator Link group to develop an interconnect standard for AI accelerator chips. An interface shared across suppliers gives system builders more ways to combine components, although a proposed standard does not by itself make every combination interchangeable. Google made the design choice more concrete in 2026: one of its TPUs planned for 2027 is intended to pair with an AMD CPU, while another uses a dual-chip package for long-context tasks. Neither has launched yet. Both show Google specifying the CPU or package around the accelerator and workload.

Modal plans to add Nvidia Vera CPUs for hosting agents alongside its existing Intel and AMD chips. That mixed estate is a useful correction to the claim that AI makes CPUs irrelevant. Providers still need them; they need freedom to select among them. Arm’s plan to sell an in-house data-center chip supplies a second reversal: a company built around licensing processor designs also sees reason to sell a finished processor.

A finished chip needs the capacity around its logic

A leading-edge die cannot ship as a useful AI system if its memory, package, or manufacturing slot is missing. TSMC and its Taiwan partners have become increasingly important to advanced packaging, while TSMC’s N3 logic-wafer capacity has itself been identified as a major AI-industry constraint. Process technology has not ceased to matter; more parts of the production route can now hold up the same delivery.

Intel said in 2023 that the United States held just 3% of global advanced-packaging capacity. A domestic logic fab therefore did not amount to a complete domestic route for advanced AI hardware. The packaging constraint has an address: the facilities that connect dies and memory closely enough for a finished system to work.

Nvidia has begun reserving far more of that wider supply chain. It said its commitments to component suppliers reached $279 billion in the second quarter of 2026, up from $119 billion in the first. Commitments are not delivered chips, but they show how much supply Nvidia sought to secure ahead of delivery. Sources also reported that Nvidia asked Samsung, SK Hynix, and Micron to use TSMC’s 3nm process for custom HBM base dies; Samsung reportedly agreed. An accelerator designer was helping specify a memory supplier’s foundry choice. In this capacity-financing market, a purchase commitment reaches well beyond the chip on the order form.

Intel can sell what it once reserved for itself

Intel’s own CPUs remain one test of its manufacturing work. Its 18A-based Xeon 6+ Clearwater Forest combines up to 288 efficiency cores with Foveros Direct 3D packaging. Intel says one Xeon 6990E+ model offers 30% better performance per thread and up to 30% greater energy efficiency than AMD’s Epyc 9965. Those are Intel’s comparisons, not independent evidence that customers will deploy it at scale. They do establish what Intel is offering: CPU design, a newer process, and advanced packaging together.

External customers impose a different test. Sources reported that Intel was in talks with Google and Amazon to provide advanced-packaging services. Another report said Google had ordered Intel manufacturing for more than three million TPUs in 2028 and that Nvidia was testing Intel technology and running 18A trials. The reported order, talks, and trials are prospective engagements, not evidence of sustained production volume. Google’s TPU would also be Google’s architecture. Intel would have to earn the manufacturing work on delivery, rather than on software compatibility.

That leaves Intel with two legitimate businesses to prove at once. A customer can choose its Xeon because the processor meets a system’s needs. Another can choose Intel’s process or packaging while buying an AMD host CPU or designing an accelerator of its own. The second customer has no reason to make Intel’s product roadmap the condition of using Intel’s factory.

Intel’s guidance put delivery ahead of possibility

Intel reported third-quarter revenue of $16.5 billion, up 21% year over year, then forecast fourth-quarter revenue and adjusted earnings per share below analyst estimates, citing continued softness in the global PC market. Its shares had closed up 23.6% on a day in August amid signs of renewed AI-driven growth; after the October guidance, they fell 28.58%. Neither trading day measured the eventual yield of 18A or the fulfillment of a 2028 TPU order. Both showed how sharply investors distinguished a possible manufacturing business from current sales and guidance.

Intel appeared in a higher share of enterprise-framed articles across the compared 2024–26 periods—22.1%, up from 16.7%—even as its total article count fell from 204 to 172. That is a measure of attention, not Intel’s enterprise revenue. The business evidence will come from shipped CPUs, completed packages, reliable foundry output, and customers willing to return with another design.

The old Intel blueprint put its name on the processor inside the machine. On the reported Google TPU work, Google brings the chip drawing to Intel’s factory gate.