Bloomberg Talks: Matt Garman (Podcast)

At re:Invent, AWS CEO Matt Garman described the company’s relationship with Anthropic as "incredibly strong," noting the startup’s enormous demand for compute has driven it to use multiple providers. Garman also highlighted AWS’s new Trainium3 AI training chip, positioning AWS to capture rising AI training workloads and underscoring continued competition among cloud providers for large-model customers.

Analysis

Market structure: Amazon (AMZN) and hyperscale cloud providers win as Anthropic’s multi-provider demand validates large-volume, high-margin GPU/accelerator consumption; Trainium3 increases AWS vertical integration and can compress competitor pricing power (Nvidia pressure at the margin but not immediate displacement). Smaller cloud vendors and GPU-resellers are losers as OEM differentiation and scale economics (power, interconnect, data‑center real estate) favor providers that own silicon + stack. Expect 6–18 month upward pressure on capital spending for data-center power/infrastructure and tighter spot supply for high-memory accelerators, supporting chip makers’ near-term pricing while increasing downward pricing leverage for commodity CPU instances.

Risk assessment: Tail risks include accelerated antitrust/regulatory scrutiny (US/EU) within 12 months, a Trainium3 performance miss vs Nvidia leading to customer churn, or Anthropic consolidating exclusively with non-AWS providers — each could knock 10–20% off implied cloud revenue growth expectations. Near-term (days–weeks) headlines from re:Invent or Anthropic benchmarks will move sentiment; medium-term (3–12 months) model-training cadence and chip delivery cycles determine revenue recognition; long-term (2–5 years) industry structure depends on adoption of custom silicon and interconnect standards. Hidden dependencies: fab capacity, power/ cooling constraints, and software stack lock-in (middleware) that governs switching costs.

Trade implications: Tactical alpha favors owning differentiated infra exposures (AMZN, NVDA) while hedging regulatory and performance risk; price shocks in options implied vol will create asymmetrical opportunities. Favor call/vertical spreads on AMZN to capture structural cloud + Trainium uptake over 6–12 months; consider pair trades long AMZN vs short legacy cloud/ERP providers whose AI TAM is limited. Monitor spot rental market for accelerators and second‑hand GPU prices as a real-time signal of training demand.

Key Decisions for Investors

• Establish a 2–3% portfolio long position in AMZN via a 6–12 month 15–25% OTM call spread (size to equal 2–3% equity exposure). Target +30–60% upside if Trainium3 adoption accelerates; set a hard stop-loss at 10% mark-to-market loss and trim 50% at +30%.
• Allocate 1.5–2% portfolio to NVDA through 9–15 month LEAP calls (buy Jan 2027 ~20% OTM or nearest available) to capture continued GPU demand; hedge 25% of notional with short-dated 30–60 day puts if implied volatility spikes above historical 60-day by >6 vol points.
• Initiate a pair trade: long AMZN (1.5% portfolio) and short ORCL (0.75% portfolio) — ORCL is chosen for weaker AI compute TAM and higher legacy revenue risk; rebalance after quarterly results or if ORCL outperforms by >8% in 30 days.
• If AMZN implied volatility ahead of re:Invent or quarterly results is < historical 90-day vol by >5 vol points, buy a 30–60 day straddle sized at 0.5% portfolio to capture binary headline risk; exit within 5 trading days of the event or if realized move <6% and time decay >50%.
• Trigger-based risk control: reduce AI/infra gross long exposure by 50% within 30 days if US/EU regulators file a formal antitrust complaint against AWS or if AWS global cloud share falls >200 basis points QoQ in reported metrics (watch quarterly earnings and third‑party share trackers).