Nvidia GPUs are everywhere. Here are the ways companies are accessing them

Nvidia shares reached another record, bringing its market capitalization close to $6 trillion, as management forecast $108 billion in October-quarter revenue, up 89% year over year. GPU demand is driving a rapid expansion in supply options: SemiAnalysis counted 323 Nvidia GPU providers in September, versus 209 less than 11 months earlier, while the hourly spot price for an Nvidia B200 GPU has more than doubled since March. The article highlights capacity constraints and varied access models across hyperscalers, neoclouds, customer-owned hardware and on-premises data centers.

Analysis

The key shift is that GPU scarcity is creating a market for compute access, not just chips. That broadens NVDA’s route to end demand, but also moves bargaining power toward whoever can secure power, data-center space, financing, and reliable delivery. Hyperscalers retain an enterprise-trust and full-stack advantage; neoclouds can monetize unmet demand, but their returns are more sensitive to utilization, deployment delays, and the cost of financing capacity. More providers do not necessarily mean durable pricing power: fragmented supply can make buyers more willing to switch, while spot-rate declines would expose operators that committed capital before demand arrived.

Over days, scarcity narratives can support NVDA and listed GPU-cloud names. Over 1–3 months, watch GPU delivery schedules, realized rental rates, utilization, and customer payment terms. Over 6–18 months, customer-owned hardware and bring-your-own-GPU arrangements could reduce cloud operators’ capital burden, but also make infrastructure economics less differentiated. NVDA’s increasing customer concentration in receivables warrants monitoring—not evidence by itself of credit stress. Compare receivables growth and collection metrics with revenue before drawing that conclusion.

The contrarian risk is that investors treat current scarcity pricing as a durable margin pool. Capacity additions, weaker-than-expected utilization, or a shift to customer-owned GPUs could pressure neocloud economics before chip demand itself breaks. Conversely, persistent power and deployment bottlenecks could keep compute prices elevated even as GPU access becomes more widely distributed.

Key Decisions for Investors

• Relative-value watch: consider a small long NVDA / short CRWV pair rather than a broad bet on every compute provider. The thesis is that NVDA sells into multiple access channels, while CRWV is more exposed to the economics of bringing capacity online and keeping it utilized. Size conservatively; CRWV’s tight capacity and rising rental rates could drive sharp upside against the short.
• Do not add to the pair solely on headline demand. Before entry, check relative valuation, CRWV’s funding and capacity-delivery disclosures, and NVDA receivables/collection trends. Falsify the relative thesis if CRWV brings capacity online on schedule with sustained utilization and rental pricing, or if NVDA’s reported demand indicators weaken.
• Treat NBIS and other neocloud exposure as catalyst-sensitive, not as a blanket sector overweight: track customer commitments versus deployed capacity, power availability, and realized pricing over the next 1–3 months. Delayed installations or falling spot rates are warning signals.
• Monitor ORCL’s bring-your-own-hardware model as a possible capital-efficiency catalyst, but wait for evidence on customer adoption and economics; company claims about improved returns are not yet proof of material earnings impact. A broader 6–18 month watch item is whether enterprise on-premises deployments displace rented workloads.