Teardown of an AI Server — and the Few Names Behind Every Layer
An AI server is the most fought-over machine on Earth right now. Strip one down to the silicon and the same pattern shows up at almost every layer: the whole industry funnels through one or two names. We didn’t decide that — the companies did, and they wrote it down. Every link below is a verbatim sentence from a public SEC filing that names the counterparty, one click from the source. We start with the order that sets everything in motion, and end at the machine that tests the last chip.
1 · The order — six gigawatts before a single chip exists
Every teardown should start before the machine exists — with the purchase order. AMD's annual report discloses a product-purchase agreement with OpenAI to deploy six gigawatts of AMD GPUs. Demand at that scale bends the whole industry around it: NVIDIA paid $5.0 billion for a stake in Intel, and Cadence co-developed an AI supercomputer with NVIDIA.
“In October 2025, we entered into a product purchase agreement with OpenAI OpCo, LLC, (OpenAI) to deploy 6 gigawatts of AMD GPUs”— AMD 10-K, filed 2026-02-04 · View source filing ↗
“we entered into an agreement with NVIDIA to issue and sell to NVIDIA 215 million shares of our common stock at $23.28 per share for an aggregate cash purchase price of $5.0 billion.”— INTC 10-K, filed 2026-01-23 · View source filing ↗
“In May 2025, we announced the M2000 supercomputer. The M2000 is a specialized, AI-accelerated platform co-developed with NVIDIA to dramatically speed up complex simulations in fields like semiconductor design, aerospace and defense, and drug discovery.”— CDNS 10-K, filed 2026-02-19 · View source filing ↗
2 · The die — only a few fabs on Earth can print it
An accelerator begins life as a blank wafer, and only a handful of fabs can print one at the leading edge. NVIDIA doesn't keep it quiet — it names its foundries outright, TSMC and Samsung, in its own annual report. AMD names GlobalFoundries for wafers at its larger process nodes. There is no long tail at this layer; the concentration is the story.
“We utilize foundries, such as Taiwan Semiconductor Manufacturing Company Limited, or TSMC, and Samsung Electronics Co., Ltd., or Samsung, to produce our semiconductor wafers.”— NVDA 10-K, filed 2026-02-25 · View source filing ↗
▸ Samsung → NVIDIA — named in the same filing sentence shown above · View source ↗
“We rely primarily on GLOBALFOUNDRIES Inc. (GF) for wafers for microprocessor and GPU products manufactured at process nodes larger than 7 nm.”— AMD 10-K, filed 2026-02-04 · View source filing ↗
3 · The memory — three names, and NVIDIA lists all three
Right beside the compute sits the memory — and this is where the AI boom actually bottlenecks. NVIDIA names exactly three memory suppliers — SK Hynix, Micron and Samsung — in a single sentence of its annual report. When you hear that memory is “sold out through next year,” this is the sentence that tells you whose year it is.
In teardown terms this layer is HBM — high-bandwidth memory stacked beside the GPU. The filings quoted below say only “memory”; we show their exact words, not ours.
“We purchase memory from SK Hynix Inc., Micron Technology, Inc., and Samsung.”— NVDA 10-K, filed 2026-02-25 · View source filing ↗
▸ Micron → NVIDIA — named in the same filing sentence shown above · View source ↗
4 · The package — ask three chipmakers, hear the same names
Packaging used to be the layer nobody mentioned; now it decides how many accelerators exist. Ask different chip companies who assembles and tests their silicon, and the filings answer with the same short list: Broadcom's single sentence names five contract manufacturers — TSMC, ASE, Foxconn, Amkor and SPIL; Astera Labs names ASE and Amkor; AMD points to its packaging JV with SPIL. ASE itself discloses a strategic alliance with TSMC dating back to 1997.
In teardown terms this is advanced packaging — “CoWoS” is the word you've heard. The filings below say “assemble, package and test”; the quote cards show their words verbatim.
“We use Advanced Semiconductor Engineering and Amkor Technologies to assemble, package, and test our ICs.”— ALAB 10-K, filed 2026-02-20 · View source filing ↗
“We use third-party CMs for a significant majority of our assembly and test operations, including TSMC, Advanced Semiconductor Engineering, Inc., Foxconn Technology Group, Amkor Technology, Inc. and Siliconware Precision Industries Co., Ltd.”— AVGO 10-K, filed 2025-12-18 · View source filing ↗
“Since 1997, we have maintained a strategic alliance with TSMC, which designates us as their non-exclusive preferred provider of packaging and testing services for semiconductors manufactured by TSMC.”— ASX 20-F, filed 2026-04-01 · View source filing ↗
“The ATMP JVs, Siliconware Precision Industries Ltd. (SPIL) and King Yuan Electronics Company (KYEC) provide ATMP services for our products.”— AMD 10-K, filed 2026-02-04 · View source filing ↗
5 · The network — “predominant” is the filing's own word
A rack of GPUs is only as useful as the fabric between them — and this layer holds the strongest single-vendor sentence in our corpus: Arista tells investors it is “primarily reliant” on its “predominant merchant silicon vendor,” Broadcom. The dependency runs down the stack: Credo fabbed its wafers exclusively at TSMC and issued an investment warrant to Amazon; Astera Labs fabricates all of its ICs at TSMC; NVIDIA alone was 27.6% of Fabrinet's revenue; and UMC licensed a co-packaged-optics process from imec.
“we are primarily reliant upon our predominant merchant silicon vendor, Broadcom, for our switching chips.”— ANET 10-K, filed 2026-02-17 · View source filing ↗
“In fiscal year 2026, we exclusively used Taiwan Semiconductor Manufacturing Company Limited (TSMC) for semiconductor wafer production.”— CRDO 10-K, filed 2026-06-15 · View source filing ↗
“During fiscal year 2022, the Company issued a warrant to Amazon.com NV Investment Holdings LLC (Holder) to purchase an aggregate of up to 4.1 million of our ordinary shares at an exercise price of $10.74 per share (the Customer Warrant).”— CRDO 10-K, filed 2026-06-15 · View source filing ↗
“We use a fabless manufacturing model and partner with TSMC to fabricate all of our ICs.”— ALAB 10-K, filed 2026-02-20 · View source filing ↗
“We engage with independent subcontractors and contract manufacturers such as Hon Hai Precision Industry Co., Ltd., Wistron Corporation, and Fabrinet to perform assembly, testing and packaging of our final products.”— NVDA 10-K, filed 2026-02-25 · View source filing ↗
“On December 8, 2025, we announced a licensing agreement with imec, a world-leading research and innovation hub in advanced semiconductor technologies, for the transfer of imec’s iSiPP300 silicon photonics process, featuring co-packaged optics (CPO) compatibility, to accelerate our silicon photonics roadmap.”— UMC 20-F, filed 2026-04-30 · View source filing ↗
6 · The power — the layer that rarely names names
Every layer so far named its suppliers in print. Power is where the filings go quiet — most power-chip makers in our corpus never say which server builds carry their parts, so this layer is deliberately thin and we won't pad it. The exception is Alpha & Omega, which lists ODM customers including Foxconn, Quanta and Wistron, and OEM relationships including Dell, in plain text. On the GaN side, Navitas discloses a long-term strategic partnership with GlobalFoundries.
“Our ODM customers include Compal Electronics, Inc., Foxconn, Quanta Computer Incorporated, Wistron Corporation and Delta Electronics.”— AOSL 10-K, filed 2025-08-28 · View source filing ↗
▸ Alpha & Omega → Quanta Computer — named in the same filing sentence shown above · View source ↗
▸ Alpha & Omega → Wistron — named in the same filing sentence shown above · View source ↗
“We have established direct relationships with key OEMs, including Dell Inc., Hewlett-Packard Company, Samsung Group, and Stanley Black & Decker, Inc., most of whom we serve through our distributors and ODMs.”— AOSL 10-K, filed 2025-08-28 · View source filing ↗
“in November 2025, we announced a long-term strategic partnership with GlobalFoundries to develop and deliver advanced GaN solutions for critical applications in high power markets”— NVTS 10-K, filed 2026-02-27 · View source filing ↗
7 · The test — the last gate, and the testers buying into it
No accelerator ships untested, and ASE — the packaging giant from chapter 4 — names the companies whose machines do it: “primarily Teradyne, Inc., Tokyo Electron Limited, and Advantest Corporation.” The testers, in turn, are buying deeper into the layer: Teradyne's strategic partnership with Technoprobe included acquiring 10% of its equity, and Advantest bought FormFactor shares in a private placement.
“We purchase testers from major international manufacturers, primarily Teradyne, Inc., Tokyo Electron Limited, and Advantest Corporation.”— ASX 20-F, filed 2026-04-01 · View source filing ↗
▸ Teradyne → ASE — named in the same filing sentence shown above · View source ↗
“in May 2024, we closed on our strategic partnership agreement with Technoprobe which included our acquisition of 10% of the equity in Technoprobe.”— TER 10-K, filed 2026-02-19 · View source filing ↗
“On January 10, 2025, Advantest America, Inc., a Delaware corporation, acquired 334,971 shares of our common stock in a private placement for $ 44.78 per share, representing the 5-day trailing volume-weighted average price prior to signing the related private placement agreement.”— FORM 10-K, filed 2026-02-20 · View source filing ↗
Out of current scope — substrate & cooling
An AI server can’t run without its ABF substrate or its cooling — but our corpus (US 10-K / 20-F annual filings, in English) holds no sentence that names a substrate or cooling supplier. Our rule is that every relationship must be backed by a filing that names the counterparty, so we leave these layers blank rather than invent them; they’ll appear as the source data expands.
Where the trail ends: the graph has no filing that says a GPU maker sells to a specific server brand, so this teardown stops where the sourced evidence does — at the contract manufacturers (Foxconn, Wistron) that assemble NVIDIA’s products.