Vultr to invest $1 billion in Ohio AI cluster using AMD chips

Vultr will invest more than $1 billion to build a 50-megawatt, 24,000-chip AI cluster in Springfield, Ohio, deploying AMD Instinct MI355X GPUs with an ethernet fabric and expects the cluster online by early 2026. The company — which recorded AI infrastructure as the majority of its revenue over the past two years and secured a $3.5 billion valuation in December 2024 — is financing the project via an extended line of credit from a bank syndicate including BofA, JPMorgan, Goldman Sachs and Wells Fargo. Vultr positions itself as a lower-cost alternative to hyperscalers, expects to pre-sell the cluster, and already has MI355X hardware in Chicago for customer testing, indicating near-term commercial traction for its AMD-powered offering.

Analysis

Market structure: Vultr’s 24,000‑chip, 50 MW buildout (online early 2026) materially increases rentable AMD GPU capacity and is an explicit low‑cost alternative to hyperscalers. Direct winners are AMD (chip sales, validation via AMD Ventures) and Vultr (volume-driven revenue growth); incumbent hyperscalers face margin pressure on commodity AI workloads and potential 10–25% downward pressure on spot GPU instance rents over 12–18 months. Secondary beneficiaries include regional colo operators and banks underwriting the debt (BAC/JPM/GS/WFC) via fees.

Risk assessment: Tail risks include project capex/operational delays, power or interconnect bottlenecks (ethernet fabric limits vs NVLink for large‑model training), and a financing shock if credit costs rise >200bps or banks tighten lines. Immediate (days) effects: AMD share reaction and bank credit spreads; short term (weeks–months): customer pilots and announced pre‑sales; long term (≥12 months): durable supply increase that could compress ASPs for premium GPUs. Hidden dependency: many enterprise customers will still prefer NVLink/HBM architectures for very large models, capping Vultr’s TAM for top‑end training.

Trade implications: Tactical: express AMD exposure with limited downside via 12–18 month call spreads (capture device demand through 2026). Rotate modestly into bank lenders (BAC/JPM) to capture financing fees and floating‑rate benefit (1–2% overweight for 6–12 months). Avoid outright large NVDA short; instead use small relative trades (long AMD vs short NVDA skewed 2:1) to play potential share shifts while hedging market beta.

Contrarian angle: Consensus underestimates performance limits of ethernet‑based interconnect for large model training—Vultr may attract inference/medium‑scale training but not top hyperscale workloads, muting long‑term GPU ASP gains. The market may be underpricing bank underwriting risk if macro tightens; conversely a faster ramp of lower‑cost GPU supply could unexpectedly compress NVDA multiples. Monitor GPU spot rents and pre‑sale contract announcements as leading indicators of durable demand.

Key Decisions for Investors

• Establish a 2% portfolio position in AMD (AMD) via a 12–18 month call spread: buy ATM Jan/Jun 2027 calls and sell calls ~15% above ATM to cap cost. Target asymmetric return of +30–50% if cluster rollouts boost AMD server GPU ASPs; cut if AMD falls 20% from entry or implied vol rises materially.
• Overweight Bank of America (BAC) and JPMorgan (JPM) by 1% each for 6–12 months to capture financing fees and floating‑rate tailwinds from syndicated credit; trim positions if bank CDS widens >75bps or NIM compresses by >50bps.
• Execute a small relative‑value trade: long AMD (1.5% notional) vs short NVDA (0.75% notional) executed within 2–6 weeks to play share rotation into AMD‑powered, price‑sensitive cloud providers. Hedge market beta with index futures to keep net directional exposure limited.
• Set active triggers: if public GPU spot rental indices fall >15% within 6 months or Vultr fails to announce ≥50% pre‑sales for the Ohio cluster by mid‑2025, initiate a 0.5–1.0% protective short on NVDA via buying Dec‑2026 puts (to hedge downside from ASP compression).