AWS CEO: AI Compute Demand So Strong No A100 Server Has Ever Been Retired

What

Why it matters

Open questions

• TSMC is reportedly working to double overall CoWoS capacity[22][23][24], but does the expansion specifically relieve the CoWoS-L bottleneck constraining Rubin Ultra[21] on a timeline that matters for that platform's competitive window?

• Nvidia's 16-high HBM request targets Q4 2026[16][17] — does this trigger a redesign cycle at all three suppliers analogous to the Rubin spec changes that drove 2026's HBM4 delays, or can suppliers adapt within current production architectures?

• With all three suppliers in active HBM4 production[14][36], does the contested Samsung-Micron #2 HBM ranking[27][28] now turn primarily on Nvidia qualification milestones and allocation decisions rather than production readiness?

• If Rubin Ultra's CoWoS-L delay[21] extends into 2027, does base Rubin[20] already lock in enough of Nvidia's next-gen installed base to limit AMD MI500's 2H 2027 competitive window[25], or does the gap create a real switching opportunity?

Narrative

Timeline

• 2025-10-30: Samsung's entire 2026 HBM supply allocation sells out after beginning Nvidia HBM shipments in Q3 2025 — the earliest documented HBM demand saturation anchor in the thread. [13]
• 2026-01-01: TrendForce publishes 2026 HBM Outlook framed as 'HBM4 Delays and HBM3e Dominance'; multiple industry analyses report TSMC working to double CoWoS packaging capacity through 2026 to relieve AI GPU demand. [10][22][23][24]
• 2026-01-05: AWS raises EC2 Capacity Block prices 15% in a uniform ML pricing adjustment. [44][45]
• 2026-01-07: AMD confirms Instinct MI500 for 2H 2027 on 2nm CDNA 6 and MI400 for 2026 at CES; The Register unpacks AMD's full datacenter GPU roadmap. [25][26][46]
• 2026-01-15: Wedbush publishes analysis formalizing the Nvidia Rubin architecture-to-HBM4-delay causal chain at Samsung and SK Hynix. [19]
• 2026-02-12: Samsung claims commercial HBM4 shipments to Nvidia began and is near a deal to supply over 30% of Nvidia's HBM4 needs. [36][47]
• 2026-04-09: Jassy's 2025 shareholder letter: AWS chips at $20B+ annual run rate; Trainium2 largely sold out, Trainium3 nearly fully subscribed, Trainium4 reserved 18 months ahead; Amazon considering external Trainium sales at potentially $50B. [3][4][39]
• 2026-04-24: Intel Q1 FY2026 earnings confirm agentic AI demand driving CPU revenue; Q2 guidance raised; Intel exec warns CPU shortage hitting 'everyone.' [40][41][48]
• 2026-04-25: Intel stock surges to all-time high of $95.73, the highest forward P/E of any large-cap chip stock. [29][30]
• 2026-04-26: AWS CEO Matt Garman publicly states AWS has never retired a single Nvidia A100 server and is completely sold out, citing demand as 'almost insatiable.' [1][2]
• 2026-04-27: SemiAnalysis H100 one-year rental price index documents nearly 40% surge over six months; Azure confirms older VM retirements while A100-based instances remain active. [5][49]
• 2026-04-30: Amazon tripled its CPU server count and still ran out of capacity; Meta signs multibillion-dollar deal to deploy tens of millions of AWS Graviton5 cores for agentic AI. [38][50]
• 2026-05-02: AMD Q1 2026 earnings beat; Lisa Su states server CPU demand 'far exceeded expectations, supply now tightening'; Intel and AMD confirm approximately 15% server CPU price increases — Intel's third round. [9][7][8]
• 2026-05-03: Samsung and SK Hynix issue public warnings that AI-driven memory shortages could persist until 2027 and beyond; SupplyFrame Intelligence extends the outlook to 2030. [11][12][51]
• 2026-05-20: Computex 2026: Micron announces high-volume HBM4 production for Nvidia Vera Rubin; Nvidia requests 16-high HBM stacks from suppliers by Q4 2026; Blackwell GPU rental reaches $4.08/hr (+48%); SK Hynix confirmed at 62% HBM market share. [14][18][16][6][27]
• 2026-05-25: Rubin Ultra identified as specifically facing CoWoS-L advanced packaging bottleneck at TSMC; Samsung reportedly reclaims #2 HBM share from Micron; Micron Q2 2026 record revenues, stock up 120% YTD. [21][28][34][35]

Perspectives

Tensions

• Samsung vs. Micron for the #2 HBM market share position: contested across three signals — one analyst placed Micron ahead[27], a report indicates Samsung reclaimed second[28], and Micron's Computex HV HBM4 production[14] may shift the picture again. [27][28][14]
• CoWoS capacity expansion vs. Rubin Ultra bottleneck timing: TSMC is reportedly doubling overall CoWoS capacity[22][23], but whether that expansion resolves the CoWoS-L constraint for Rubin Ultra specifically — and arrives in time to matter for that platform — remains open[21]. [22][23][24][21]
• Base Rubin vs. Rubin Ultra: base Rubin in full production since CES 2026[20], while Rubin Ultra faces a specific CoWoS-L bottleneck at TSMC[21] — a two-track timeline within the same architecture family with distinct competitive implications for AMD and memory suppliers. [20][21][19]
• AMD MI500 timing: TechRadar asks whether MI500 arrives 'too late' given Nvidia's Vera Rubin 2026 target[42], while Rubin Ultra's CoWoS-L delay[21] could make AMD's 2H 2027 window more competitive — with base Rubin's head start as the countervailing factor[20]. [42][21][25][20]
• Is AMD's record CPU market share gain supply-driven (temporary) or architectural preference (durable)? Multiple researchers attribute AMD's gains primarily to Intel supply constraints[31][32][33], but AMD's own supply tightening[9] complicates indefinite share-taking. [31][32][33][9]
• Intel's highest forward P/E of any large-cap chip stock[30] versus AMD's record market share gains attributed to Intel supply failures[43]: Intel's premium valuation prices in a recovery that AMD market share data suggests may not fully materialize. [30][29][43][31]