chokepoints.ai
SUBSCRIBE
10 layers580 nodes2,376 dependencies9 chokepoints112 bottlenecks6,500+ companiesnode size = companies identified

Servers, systems, accelerator boards, and racks

BOTTLENECK

Three ODMs capture GB200 NVL72 assembly margin, and thermal density prevents rapid qualification of a second source.

Server and rack-level systems integrating accelerators, CPUs, memory, and networking. Physical density, thermal limits, and interconnect topology determine how AI capacity deploys. ODMs (Quanta, Foxconn, Wistron) capture assembly margin on GB200 NVL72 racks; hyperscalers are the primary buyers.

The hardware assembly layer that integrates accelerators, CPUs, memory, storage, and networking into complete servers and rack-level systems; spans from the GPU baseboard through branded OEM servers, ODM white-box servers, and rack integration.

What the evidence shows

Server hardware options include offerings from over a dozen vendors.

techtarget.com

AI servers with accelerators accounted for over 70% of server revenue in Q1 2026.

tomshardware.com
RESCORED JUL 2026consolidatedscaling162 companies

Who supplies it

The global AI server assembly base is concentrated in Taiwan, which accounts for about 90% of worldwide AI server production by MIC and the Ministry of Economic Affairs. That geographic concentration is amplified by Nvidia, which reportedly booked entire server plant capacity through 2026 for Blackwell and Rubin systems and pushed out other potential customers. Hyperscalers and large cloud providers act as the primary off-takers of the resulting racks, with one customer receiving 100,000 Nvidia GPUs from a single Dell factory in six weeks.[10][3][5]

System integrators that brand and configure accelerators include Dell, HPE, Lenovo, Supermicro, and IBM, with rack-scale offerings led by Dell, Supermicro, and HPE. Supermicro alone reports the ability to build 4,000 racks per month, scaling to 5,000 per month by its fiscal June 2024 quarter, including 1,500 liquid-cooled racks, and posted fiscal Q2 2024 revenue of $3.67 billion. Rack-scale AI made up 48.6% of Supermicro's trailing twelve-month sales, or $4.5 billion of $9.25 billion.[1][14][8]

Component suppliers feed the integrators and concentrate around a few firms. Samsung Electro-Mechanics signed a 1.5 trillion won long-term silicon-capacitor supply agreement in May 2026 and won separate MLCC contracts for AI servers of 454 billion won in June 2026 and 295.1 billion won in July 2026. TrendForce reported lead times of 35 to 40 weeks for power management chips and 21 to 26 weeks for baseboard management controllers in general servers, with overall 2026 server shipment growth downgraded from 20% to 13%.[11][4]

Who controls it

DellHewlett Packard EnterpriseLenovo+159 more tracked
$145B market · 20257.2% CAGRsource

Where it sits in the stack

Takes in: GPU modules, CPUs, NICs, memory, storage, chassis, PSUs, networking hardware

Sends on: Complete AI servers, rack systems, and turnkey AI infrastructure

depends onMerchant GPU and general-purpose AI acceleratorsHyperscaler custom ASICs and XPUsHost server CPUsNetwork interface cards, SmartNICs, and DPUsMemory subsystemsStorage subsystemsNetworking and interconnect hardwareIn-rack and on-board power delivery (IT side)AI accelerators and silicon
feeds intoLand, siting and entitlementFacility ownership and operationWhite-space fit-outCloud compute platformsHyperscale AI cloud platformsNeocloud / GPU-specialist cloud platformsBare-metal and GPU-as-a-service (dedicated)GPU broker / aggregator marketplacesVirtualisation and multi-tenancyFrontier closed-model developersSupply-chain management and procurement services
view in atlas

What would break it

Component bottlenecks are widening elsewhere in the supply chain. Power management chip lead times run 35 to 40 weeks and baseboard management controller lead times 21 to 26 weeks for general-purpose products, driving TrendForce to cut its 2026 server shipment forecast from 20% to 13%. The accelerator market itself is fragmenting as AWS designs its own AI chips and AMD and Qualcomm grow, so rack integrators must qualify multiple chip families rather than optimizing around one supplier.[4][2]

What to watch

Several rack-scale platforms are scheduled to ship in the second half of 2025 and 2026. The flagship PowerEdge XE8712 with 144 B200 GPUs and hybrid DLC cooling is available from H2 2025, the PowerEdge XE9680L with 8 B200 GPUs and direct liquid cooling is priced at $500K to $600K for H2 2025, and 18th-generation PowerEdge models on AMD EPYC Venice with Nvidia Rubin are expected to ship in H2 2026. AMD's Helios rack with MI400 GPUs and EPYC Venice is due in 2026, and Supermicro's rack build rate is targeted to reach 5,000 per month by its fiscal June 2024 quarter.[6][7][13][14]

Back-end packaging and memory capacity additions stretch into 2027 and 2028 and will shape how many accelerators can be deployed. TSMC's CoWoS packaging is forecast to reach 140,000 units per month by 2026 and 190,000 to 200,000 units per month by 2027, against Nvidia demand of 630,000 CoWoS units in 2026 and 1,005,000 in 2027. Amkor's new Arizona packaging facility is not expected to begin production until 2028, and Micron's newest major memory fab in Singapore is scheduled to begin production in the second half of 2028.[9][12]

Related nodes

AI accelerators and siliconHost server CPUsNetwork interface cards, SmartNICs, and DPUsMemory subsystemsStorage subsystemsNetworking and interconnect hardware

Sources

  1. intuitionlabs.ai · 2026-03-01T06:03:17
  2. crn.com · 2026-03-16T00:00:00
  3. tomshardware.com · 2025-06-21
  4. theregister.com · 2026-04-23
  5. blogs.nvidia.com · May 27, 2025
  6. slyd.com · H2 2025
  7. vrlatech.com
  8. alliantglobal.com · 2028
  9. winbuzzer.com · July 4, 2026
  10. penchan.co · 2025 H2
  11. asiae.co.kr · 27 Jul.2026
  12. randtech.com
  13. servethehome.com · June 12, 2025
  14. nextplatform.com · 2024-01-30

Full scorecard, owner shares, supply edges and the full tracked roster are in the desk letter.

GET THE BRIEFING