Nvidia and AMD Accelerators: Roadmaps, Pricing, Supply — 2026-09-24
This week's headline news: GPU consumer-level supply is collapsing as AI demand pulls dies away from retail — the RTX 5090 hit $9,500 with US stock vanishing, and AMD raised prices across its lineup. AMD crossed $1 trillion in market cap on Sept 21, while Samsung ramped HBM4-related wafer carrier capacity for Nvidia's Rubin era. China's domestic AI accelerator makers hiked prices 20–50% amid the same HBM shortage.
Nvidia and AMD Accelerators: Roadmaps, Pricing, Supply — 2026-09-24
Top developments
RTX 5090 vanishes from US retail at $9,500 as Nvidia shifts dies to Pro cards
Published 2026-09-23, the report says the RTX 5090 price has reached $9,500 as US retail stock disappears, and that Nvidia is reportedly shifting GPU dies to AI-focused RTX Pro cards, per Tom's Hardware and PC Gamer. This signals extreme Die allocation pressure at the low end of Blackwell as data-centre demand dominates. It underscores why Blackwell volumes are tapering while the Rubin transition becomes Nvidia's priority.
AMD crosses $1 trillion market cap on MI400 momentum
AMD hit a $1 trillion market cap on Sept 21, 2026, joining Nvidia, Broadcom and SK hynix, after a ~10% rally that put it $352B past Intel. Korean media (MS TODAY, EBN) framed the milestone as evidence AI competition is shifting from individual GPU performance to full data-centre systems — directly relevant to AMD's MI450/Helios platform challenge to Nvidia's Vera Rubin roadmap.
AMD raises AI chip, Radeon and chipset prices 10% for Q4
AMD has warned partners of a 10% price hike on AI chips (sparing Ryzen), effective Q4 2026, citing TSMC wafer costs; Korean local media also reported a broad 10% increase covering GPUs, motherboards and CPUs. Intel separately announced a ~10% CPU price increase from Oct 5, showing the shortage is industry-wide. This directly affects expected Instinct MI350/MI400 system pricing and data-centre budgets.
Samsung expands HBM4 carrier capacity 2.5x for Nvidia's Rubin
On Sept 21–22, Wccftech reported that Samsung is increasing silicon/glass-substrate carrier procurement for HBM4 by roughly 2.5x amid surging AI memory demand, as covered by Korea's Global Economic. The move signals the supply-side race to support the Vera Rubin platform, which Nvidia targets for fall 2026 shipment.
China's domestic AI chipmakers hike prices 20–50% on HBM shortage
Huawei, Cambricon, MetaX and Iluvatar CoreX raised AI accelerator prices 20–50% as HBM scarcity squeezes China's domestic accelerator makers, per a Sept 22 report. This matters for Nvidia's China-specific strategy: domestic rivals are losing their cost advantage, potentially widening the window for whatever compliant export variants Nvidia can ship.
Local view
Korean media are heavily focused on the memory supply chain and AMD's rise. 중앙이코노미뉴스 reported AMD is pitching its next-gen server CPU "EPYC 9006 Venice" with claimed ~2.24x performance advantage over Nvidia's Vera CPU, and flagged Samsung and SK hynix server DRAM as stocks to watch.
테크M reported that Akamai is preparing GPU deployments on Nvidia Blackwell Ultra and Vera Rubin, aiming to expand its AI business from inference to training customers via a globally distributed "AI Grid".
Sina's Chinese-language chip digest (Sept 24) flagged TSMC's new pricing round, reported by Digitimes, with hikes planned from January 2027 — another forward cost pressure on both Nvidia and AMD accelerators.
Context & numbers
- JPMorgan forecasts Rubin shipments reaching roughly 5.7 million GPUs in 2026 as Blackwell volumes taper to about 1.8 million, per a Sept 22 Tech Insider Canada report on AMD's 432GB HBM4 MI400 challenge to Nvidia.
- Full rack pricing from recently updated guides: a GB200 NVL72 rack runs roughly $2M–$3M (~$27,800–$41,700 per GPU), while the GB300 NVL72 sits in the $3M–$4M range.
- The H200 rents from roughly $2.71/GPU-hour on marketplaces up to $27.04/GPU-hour on constrained hyperscaler allocations; purchase sits near ~$30,000.
- China's liquid-cooled data-centre market reached ¥15.98B in 2025, up 45.2% YoY, as GB300 rack power levels exceed what air cooling can handle; GB300 racks are expected to ship in volume from 2026 (Sina Weibo, Sept 23).
- HBM accounts for ~48% of the B200's build cost, keeping effective B200 pricing firm near $40,000 (siliconanalysts.com; reference guide pricing, not confirmed current quotes).
On the radar
- Singapore-based Aolani announced a plan to deploy 22,000 Nvidia Blackwell Ultra GPUs across two AI factories in Malaysia and the Philippines by early 2027 (TradingKey, Sept 21) — watch for Southeast Asia capacity announcements.
- TSMC's reported 2027 price hikes (Digitimes, via Sina) could feed into further AI accelerator price increases — a rumor flagged as unconfirmed by TSMC.
- DRAM price dynamics are reportedly pushing next-gen consumer GPU launches (RTX 60 / RDNA 5) toward 2028; per Tech Insider's ~1-week-old report, "GPU relief" is now not expected before then.
- Anthropic is reportedly in early talks with Stream Data Centers to directly lease up to 1GW of data-centre capacity — early-stage and may not conclude (The Information via ChainCatcher, Sept 23).
This content was collected, curated, and summarized entirely by AI — including how and what to gather. It may contain inaccuracies. Crew does not guarantee the accuracy of any information presented here. Always verify facts on your own before acting on them. Crew assumes no legal liability for any consequences arising from reliance on this content.
Powered by
