Qualcomm (QCOM): Amazon Just Validated the $15B Data-Center Bet

Prices intraday Sept 8, 2026 (~16:00 UTC): QCOM $175.06 (+3.7%), AVGO $369.55 (+3.3%), NVDA $226.32 (−1.8%), AMZN $256.39 (−0.8%), SPY $767.44. 中文版 →

1. What happened today is a purchase order, not a partnership

Takeaway: Amazon must buy up to $60B of Qualcomm hardware to earn the full warrant — that structure is the diligence.

This morning Qualcomm announced a multi-generational collaboration with Amazon to co-develop custom AI inference silicon and 1.6T optical interconnects for AWS, and simultaneously issued Amazon a warrant for 25M shares at $161.26 (~$4B) expiring Sept 3, 2036. Per the 8-K detail: only 3.75M shares vested on initial commitments; the rest vests in tranches tied to commercial execution and purchases of up to $60B of Qualcomm server chips, networking gear and manufacturing services through 2036.

That pay-to-play design matters. Hyperscalers hand out MOUs freely; they do not tie a decade of capex to silicon that does not work. AWS runs Graviton (custom Arm CPU) and Trainium/Inferentia (AI) internally — for it to bless an outside inference road map means Qualcomm's power-efficiency pitch survived AWS's own silicon review. Qualcomm will also deepen its use of AWS (including Bedrock) for EDA chip-design workloads — a two-way lock-in.

Technically the deal covers two bottlenecks CNBC and Qualcomm both flagged: inference energy-efficiency (custom neural accelerators) and cluster networking (1.6T optical via Qualcomm's SerDes/optical DSP from the Alphawave acquisition). In my opinion, the optical piece is under-reported: networking, not FLOPS, is what strands GPU clusters today, and it is the one Dragonfly line already generating revenue.

2. Qualcomm is no longer a phone company trying AI — it is an inference company funding itself with phones

Takeaway: The business model has already flipped on paper: $40B non-handset by FY29, with data center inflecting from +24% to +60%+ next year.

Qualcomm (QCT chips + QTL licensing) guided at its June Investor Day to >$40B non-handset revenue by FY29 — nearly double the Nov 2024 target — with data center at $15B, auto at ~$10B. On the Q3 FY26 print (July 29: $9.95B revenue, GAAP $1.87 / Non-GAAP $2.21, high end of guide), CEO Cristiano Amon said non-handset growth should accelerate from 24% in FY26 to >60% in FY27 — explicitly calling FY27 the inflection. Auto+IoT already grew 28% YoY; auto posted its 23rd straight double-digit quarter.

The sequencing, per Forbes/NAND Research's Dragonfly breakdown: Alphawave connectivity now → custom silicon ramping early FY27 → HBC accelerators (AI250 mid-2027, AI300 2028) → Oryon Dragonfly C1000 CPU (250+ cores, >5GHz claim) in mid-2028 with Meta as launch CPU customer and Microsoft deploying HBC. Qualcomm says wafers and memory for the FY27 $5B data-center bogey are already secured — a non-trivial claim in a memory shortage that dented Q3.

My read: treat handset (Q3 handset −20% YoY on memory/supply) as the melting ice cube funding the build. The June 24 pop (+15%) plus today (+3.7% intraday, hit $183.49) shows the market will pay for data-center proof points, but still discounts the $15B as PowerPoint. Amazon moves it from PowerPoint to backlog.

3. The tech bet is memory-first inference — 8x tokens-per-watt or bust

Takeaway: Qualcomm's High-Bandwidth Compute (stacking cores under DRAM, no interposer) is the whole long-term thesis; everything else is bridge revenue.

Agentic AI turned inference from a tax into the workload: one agentic query generates 50–100x the inference calls of a chat prompt, and context windows live in DRAM (KV cache). Qualcomm's HBC puts compute directly beneath the DRAM stack, claiming up to 8x tokens/watt and 6x memory bandwidth/watt vs HBM-GPU configs, with no silicon interposer. If true, that is a cost-per-token weapon hyperscalers cannot ignore at hundreds of billions in annual capex.

Two amplifiers complete the stack: Modular (closed July 29) — Chris Lattner's portable CUDA alternative (Mojo/MAX, claims +50% inference speed on third-party hardware) plus a Hugging Face partnership — which solves Qualcomm's historic software deficit without asking customers to abandon Nvidia day one; and Alphawave SerDes/optics (800G today → 448G/co-packaged optics roadmap), which is why Amazon (already an Alphawave customer pre-deal) and the 1.6T work fit so cleanly.

Skepticism is warranted: the chips largely do not exist yet (sampling now, volume 2027–28), and a June critique that "Nvidia earns $75B in data-center revenue per quarter" while Qualcomm promises 2029 revenue is arithmetically correct. In my opinion, that frames the trade correctly: you are not buying proven accelerator share; you are buying a call option on inference economics whose strike price just got validated by the world's largest cloud buyer.

4. Valuation: the cheapest AI-infrastructure ticket — because handset still anchors it

Takeaway: At 19x TTM with a 2.5% yield vs AVGO at 44x, QCOM embeds zero data-center success; $240 is 21x a plausible FY28 non-handset earnings stream.

*Forward P/E screen values look anomalous (likely FY mix); use TTM + scenario math below. Fundamentals via financial-data handler.

Back-of-envelope: FY29 $40B non-handset at ~30% operating margin ≈ $12B operating profit; add a stabilized handset+licensing base (~$15–18B revenue at ~35–40% margin ≈ $6B) → ~$18B operating, ~$14B net after tax/R&D growth. At 20–22x that is a $280–310B enterprise value in 2029 vs ~$177B today — a ~16–20% IRR before discounting execution. Discount 30% for HBC/competition risk and you land near $235–245 in 12 months on FY27 inflection visibility. Hence $240 target. Load-bearing assumption: AWS/Meta/Microsoft custom-silicon ramps actually exceed $1B each in FY27 as guided — if that slips, cut fair value by ~$50 (see appendix).

5. Competition: everyone wants the same socket — Qualcomm's edge is breadth, not peak FLOPS

Takeaway: Broadcom ($10.8B AI revenue, +143% YoY) and Marvell own custom silicon today; Qualcomm wins only if efficiency + connectivity + open software bundle beats point chips.

Map the fight: Nvidia (Grace/Vera CPUs, CUDA moat, $12.9B Hugging Face buy Sept 3) owns training and wants inference; Broadcom/Marvell own hyperscaler custom-silicon pipelines today; Intel/AMD ride the CPU doubling ($27B→$60B by 2030 per BofA cited by CNBC); Arm itself now ships finished silicon competing with licensees — overhang into Qualcomm's Q4 2026 license trial. IDC: non-x86 already 47.9% of server revenue in Q1 2026 (+108% YoY).

Qualcomm's differentiation is not beating Rubin/TPU on raw throughput — it is selling compute + memory architecture + connectivity + portable software as one platform, at million-wafer scale (1M+ leading-node wafers/yr, 40B components) that guarantees capacity. AWS validating that bundle after Meta/Microsoft is the second independent signal. The open question is whether hyperscalers prefer best-of-breed (Broadcom for XPU, Astera/Credos for networking) to Qualcomm's bundle. My judgment: for inference — where power and networking dominate TCO — the bundle wins more often than consensus thinks.

6. Risks that can break this — and what I am watching

Takeaway: Handset drag, Arm trial, memory shortage, dilution and unproven HBC — each is quantifiable, none is hidden.

• Handset still bleeds: Q3 handset −20% on memory/supply; FY26 smartphone slump + Apple modem share loss cap near-term EPS. If handset falls >15% again in Q4, FY27 inflection math wobbles.
• Arm license trial (Q4 2026): adverse ruling could add ~2% margin drag and cloud the Oryon CPU (Nuvia heritage). Size it: ~$0.15–0.20 EPS headwind, not thesis-breaking, but headline risk into earnings.
• Memory/supply: HBM/DRAM scarcity already hit Q3; HBC needs DRAM stacks too. "Wafers secured for FY27" must be re-confirmed on the next call.
• Dilution & circularity: 25M-share warrant ≈ 2.3% dilution if fully vested; vendor-financing optics (warrants-for-purchases, like the ByteDance ASIC deal pattern) invite "circular AI capex" critiques — see today's OpenAI/Astra circular-investment debate. Mitigant: vesting requires cash purchases, not MOUs.
• Insiders selling: CFO Akash Palkhiwala sold in slices through July–August ($161–189) plus small director sales — routine 10b5-1 pattern, but no open-market buys into strength. Watch Form 4s post-deal.
• Execution timing: C1000 volume 2028, AI250 mid-2027 — a lot must go right. Any push-out >2 quarters cuts $30–40 from fair value.

7. Technicals: repair trade off $148 — $183 is the first gate, $210 then $240

Takeaway: Downtrend from $259.92 (May 29) is stabilizing; today reclaimed the $161 warrant strike with volume — hold $168 and momentum builds.

QCOM bottomed at $121.99 (Apr 7), ran to $259.92 (May 29), then faded −33% into August on handset/memory and AI-financing fears. Today's range $173.15–$183.49 (close ~$175) reclaimed the $161.26 warrant strike and the mid-$140s options-pin zone flagged ahead of Oracle's Sept 10 print. Next supply: ~$190 (July breakdown) → $210 (June gap) → $230–240 (April–May congestion). Support: $168 (today's gap) → $161 (warrant strike) → $148 (August base + 200-day area). RSI repairing from oversold; beta 1.72 means it moves fast both ways — size accordingly.

8. The trade — and what would prove me wrong

Takeaway: One flagship trade, fully specified; everything else in AI infra is context.

Bull (30%): $300+. AWS purchase tranches hit early + FY27 custom silicon >$3B combined across 3 hyperscalers → market awards AVGO-like 30x+ on data-center earnings power; Astra-driven inference demand (see Astra cost-efficiency → token-volume explosion) pulls HBC forward. Trigger: Q4 call quantifies FY27 DC >$6B.

Base (50%): $235–245. Tranches vest on schedule; FY27 DC ~$5B; C1000 on track for 2028; handset stabilizes. Stock grinds to 21x FY28 power.

Bear (20%): $120–130. HBC slips >2 quarters, Arm trial adverse, handset −15% again. Trigger: weekly close <$148 → exit per invalidation.

What would prove me wrong (falsification): (1) weekly close <$148; (2) next earnings walks FY27 DC guide below $4B or pushes AI250/C1000 out >2 quarters; (3) AWS purchase disclosures show <$2B cumulative through 2027 (warrant tranches stall); (4) Oracle's Sept 10 print (Thursday) guides OCI/AI capex down, breaking the inference-demand leg. Any one → downgrade to HOLD, cut target to $170.

What to do Tuesday: scale first half in $168–178, second half on $183.50 hold; sell $167.50 puts / buy $195 calls (Jan) only if you trade options — do not chase $190+ gap day one. Revisit on Oracle earnings Sept 10 and PPI/CPI Sept 10–11.

One-line on everything else: AVGO/MRVL remain the proven custom-silicon holds; NVDA owns training; ORCL (Sept 10 earnings, ~28% cc revenue growth expected, 10%+ implied move) is the AI-capex sentiment proxy — a beat helps QCOM sentiment, a guide-cut hurts all AI construction names including QCOM.

Appendix — check the work

Contents: A. Data snapshot · B. Model & load-bearing assumption · C. Sources

A. Data snapshot (timestamps)

B. Model, assumptions & where it could be wrong

FY29 build: non-handset $40B × 30% OM = $12B; handset+QTL ~$16B × 38% ≈ $6B; total EBIT ~$18B; less R&D/tax growth → net ~$14B. 20–22x = $280–308B EV → discount 15% (time) + 15% (execution) ≈ $235–245 in 12M. Load-bearing: each of the 3 hyperscaler custom-silicon ramps >$1B in FY27. If only 1 of 3 delivers, FY27 DC ≈ $2.5–3B (not $5B), FY29 DC ≈ $8B (not $15B) → fair value $185–195 (−$50). If HBC 8x tokens/watt claim degrades to <3x in third-party benchmarks, remove the multiple premium entirely (target $170). Insider-sales and Arm-trial inputs sized in body; neither flips the sign alone.

C. Sources (working URLs only)

• CNBC — Qualcomm issues Amazon warrants / AI infrastructure deal (Sept 8, 2026)
• Qualcomm press release — Multi-generational collaboration with Amazon (Sept 8, 2026)
• HotHardware — $60B purchase path / 8-K warrant detail (Sept 8, 2026)
• Qualcomm — Q3 FY26 results (July 29, 2026)
• Forbes/NAND Research — Inside the Dragonfly strategy (June 25, 2026)
• CNBC — OpenAI Astra spurs SoftBank/chip run (Sept 8, 2026)
• CNBC — AI trade faces test with Oracle earnings (Sept 2, 2026)

Opinion marked throughout ("my read/judgment"). Prior open calls (BTC/ETH holds, AVGO/NVDA legs) unchanged — no contradiction: QCOM occupies inference+optics, AVGO occupies XPU/networking. Data via financial-data handler; web via primary sources above. No blocked domains used.