JPMorgan: AI Server Supply Chain Enters Third Round of Price Hikes as ODMs Begin Raising Prices
Nashnova编辑部
J.P. Morgan finds the AI server supply chain has entered a third wave of price increases, with ODMs for the first time negotiating higher quotes on motherboards and systems — pricing power is migrating from upstream components to manufacturing services, and supply bottlenecks may persist through 2027.
Three waves of price hikes — what's different this time?
Wave one hit commodity components: memory, MLCCs — tiny capacitors used on circuit boards in nearly every electronic device. Wave two reached substrates, foundry services, and other key manufacturing steps.
Wave three has produced something rarely seen before: some ODMs — contract manufacturers that design and assemble finished servers for brand-name clients — are now proactively asking customers to accept higher quotes.
This means → price increases have moved from "parts cost more" to "assembly services cost more too." Pricing power is shifting upward across the entire chain.
How strong is demand, really?
J.P. Morgan's field checks show the top four cloud providers are running short on compute capacity and actively onboarding new AI customers.
Last year roughly 80% of incremental demand came from a handful of AI firms. This year hundreds of previously unknown buyers have appeared, making demand far more broad-based.
This reflects a shift: AI compute demand has spread from a few hyperscalers to the wider enterprise market — no longer a top-player-only game.
Shipment forecasts are up — so why the downside risk?
J.P. Morgan raised its general-server shipment growth forecasts to 22% this year and 25% next year, with Q3 and Q4 each expected to grow about 4% quarter-on-quarter.
But supply is flashing amber: some motherboard suppliers say Q3 shipments may decline sequentially, and one major ODM's quarterly growth expectation has been cut from double digits to the high single digits.
In plain terms = demand is surging, but supply can't keep up — and that gap is exactly what underpins the pricing power.
When will the component bottleneck ease?
The tightest spots are memory-related: DRAM, eMMC, SSDs, and NAND flash. Server CPUs are also approaching a broader shortage.
Analog ICs, MLCCs, and SP capacitors are similarly constrained. Under current demand assumptions, supply relief is not expected before 2027.
This means → if supply does improve in 2027, the chain could face spec downgrades and inventory corrections — a moment when valuation assumptions will be stress-tested.
Where do Nvidia, AMD, and Google stand on new platforms?
Nvidia's Vera Rubin saw two schedule adjustments tied to PCB issues but shows no further delays. Downstream shipments in H2 are estimated at roughly 1.5 million units; at the rack level, about 10,000 racks at 72 GPUs each imply around 700,000 GPUs consumed, with the gap explained by production-cycle timing.
AMD Helios feedback is limited. Oracle and Meta projects are reportedly moving forward; volume production may begin around Q1 next year, slightly behind Nvidia's timeline.
Google TPU upstream is expected at roughly 4.0–4.5 million units, downstream about 3.5 million. Amazon is accelerating Trainium production. Combined TPU-plus-Trainium shipments for the full year are estimated at 2.5–2.8 million units.
For investors, what is the key variable to watch?
Lenovo's infrastructure business reportedly grew nearly 100% year-on-year, with both enterprise and cloud revenue up over 90%. Component inflation and pricing premiums mean server revenue growth is expected to significantly outpace unit shipment growth.
In plain terms = more servers are shipping, but each one sells for more — revenue grows much faster than volume. That is what pricing power looks like on a P&L.
J.P. Morgan's bottom line: whether supply relief arrives on schedule is the single most important variable for validating supply-chain valuations in 2027.
Content is for reference only, not financial advice.