Five chip and silicon companies<\/a> are capturing the spend flooding into optical interconnects, custom XPU silicon, CPU licensing, PCIe fabric, and wafer-scale inference. Marvell posted $1.83 billion in data center revenue for Q1 FY2027, up 27% year over year, while Astera Labs grew 93% in the same period. Arm’s data center royalties more than doubled. Cerebras landed a $20 billion multi-year inference contract with OpenAI, plus a $1 billion working capital loan from the same customer.<\/p>\nWhat this means for your business<\/h2>\n
Where your infrastructure dollars land in the next 18 months depends on which layer of the stack you’re actually constrained by. Teams still optimizing purely around GPU availability are building on an assumption that’s already dated. The chokepoints have migrated to the connective tissue, optical links running at 800G to 1.6T speeds, PCIe retimers keeping chip-to-chip traffic alive at rack density, and CPU cores licensing their way into every custom AI chip shipping in volume. If your architecture reviews don’t surface those layers explicitly, you’re planning around last year’s bottleneck.<\/p>\n
The Arm royalty model deserves particular attention from anyone evaluating custom silicon partnerships. Arm doesn’t build chips; it collects a license fee upfront and then a per-unit royalty on every chip shipped, forever, across NVIDIA Vera, Google Axion, and Microsoft Cobalt. That’s not a vendor relationship your procurement team can renegotiate away. It’s a structural cost embedded in the unit economics of every major AI CPU now in production. The data center CPU market is projected above $100 billion by 2030, and Arm already reports over $2 billion in customer demand for its AGI CPU across FY27-FY28.<\/p>\n
The Cerebras situation reframes a question many CTOs have deferred, which is whether wafer-scale compute, a design that treats an entire silicon wafer as a single processor rather than dicing it into individual chips, belongs in your inference architecture. When OpenAI commits $20 billion and backs it with a $1 billion loan to the same vendor, that’s not a technology bet anymore; it’s a signal about where inference economics are heading. The margin compression Cerebras guided for year one matters less than the fact that the largest AI lab in the world chose this architecture for sustained workloads at scale. If your inference stack is still GPU-only, that’s the assumption worth stress-testing now rather than at your next vendor renewal.<\/p>\n
Concept deep-dive: PCIe Retimers<\/h2>\n
PCIe (Peripheral Component Interconnect Express) is the high-speed highway that lets processors talk to memory, storage, and each other inside a server. At the rack densities AI clusters require, those signals degrade over distance, like a phone call breaking up as you move further from the tower. A retimer is a chip that catches the signal, cleans it up, and re-transmits it at full strength. As racks get denser and chip-to-chip bandwidth demands increase, retimers shift from optional to mandatory, which is the core of Astera Labs’ business case.<\/p>\n
Based on reporting from 5 Under the Radar AI Chip Stocks Powering the Data Center Boom<\/a>, originally published 2026-07-11 09:10:00.<\/em><\/p>\n","protected":false},"excerpt":{"rendered":"Share with your CTO The AI infrastructure buildout has already moved past GPU allocation as the primary bottleneck. Five chip and silicon companies are capturing the spend flooding into optical interconnects, custom XPU silicon, CPU licensing, PCIe fabric, and wafer-scale inference. Marvell posted $1.83 billion in data center revenue for Q1 FY2027, up 27% year […]<\/p>\n","protected":false},"author":1,"featured_media":5070,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[147],"tags":[207],"tmauthors":[],"class_list":["post-5069","post","type-post","status-publish","format-standard","has-post-thumbnail","category-ai-infrastructure","tag-cto"],"_links":{"self":[{"href":"https:\/\/workai.tv\/news\/wp-json\/wp\/v2\/posts\/5069","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/workai.tv\/news\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/workai.tv\/news\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/workai.tv\/news\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/workai.tv\/news\/wp-json\/wp\/v2\/comments?post=5069"}],"version-history":[{"count":0,"href":"https:\/\/workai.tv\/news\/wp-json\/wp\/v2\/posts\/5069\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/workai.tv\/news\/wp-json\/wp\/v2\/media\/5070"}],"wp:attachment":[{"href":"https:\/\/workai.tv\/news\/wp-json\/wp\/v2\/media?parent=5069"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/workai.tv\/news\/wp-json\/wp\/v2\/categories?post=5069"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/workai.tv\/news\/wp-json\/wp\/v2\/tags?post=5069"},{"taxonomy":"tmauthors","embeddable":true,"href":"https:\/\/workai.tv\/news\/wp-json\/wp\/v2\/tmauthors?post=5069"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}