Enterprises put non-Nvidia chips 14 points ahead of Nvidia's next-gen GPUs on their evaluation lists

VentureBeat’s July VB Pulse survey shows enterprise AI accelerator optionality is rising: 39.4% of respondents said they’re likely to evaluate non-Nvidia accelerators over the next 12 months vs 25.3% for Nvidia’s next-generation Blackwell (a 14-point gap), while near-term platform-change urgency fell (0–3 months down 9.5 pts, to 28.8%). Production adoption expanded rather than meaningfully changing platforms—Azure in production jumped from 29.0% to 47.1% (+18.1 pts) and OpenAI in production from 40.2% to 49.4% (+9.2 pts), with uptime/reliability becoming a more important effectiveness measure (42.1% to 51.2%). Neoclouds show demand pull (48? ] intended more with neoclouds up from 33% to 38%, while named-provider production adoption rose from 1.9% to 5.9%), and open-source self-managed production stacks grew from 3.7% to 12.9%, suggesting implementation depth over just evaluation.

Analysis

The market implication is less “Nvidia loses share tomorrow” than “buyers now have leverage.” That changes pricing power, pilot length, and vendor lock-in economics: even if NVDA keeps the default slot, incremental capacity decisions are increasingly won on cost-per-useful-work and integration, which is a margin headwind for the dominant incumbent and a relative tailwind for anyone selling credible second-source silicon or vertically integrated cloud capacity. The first-order beneficiaries are hyperscalers with their own silicon roadmaps and procurement scale; the second-order beneficiaries are enterprises that can force better terms across the stack.

The near-term catalyst is not the survey itself but the next round of capex and earnings commentary from AMZN, GOOGL, MSFT, AMD and NVDA. If hyperscalers keep steering workloads toward in-house accelerators, the biggest P&L impact is on NVDA mix and ASPs, while AMD gains as the obvious external alternative that can be adopted without a full platform rewrite. CRWV and NBIS are more of a trading vehicle on “capacity scarcity” than a clean fundamental long: their upside depends on converting backlog into durable utilization at acceptable returns, and reliability/security friction can slow that conversion materially over 1-3 quarters.

Contrarian read: consensus may be overstating immediate substitution. Optionality often expands total AI spend rather than reallocating it, because teams hedge with multiple chips, more orchestration, and more governance layers. If Blackwell supply stays tight and NVDA guides to continued data-center acceleration, the short-NVDA thesis is vulnerable; if AMD/cloud design-win commentary does not improve over the next two earnings cycles, the diversification trade becomes narrative-only rather than earnings-supported.

Key Decisions for Investors

• Initiate a relative-value short NVDA / long AMD pair on strength, targeting a 1-3 month window; thesis is multiple compression on NVDA from procurement leverage versus improving optionality value for AMD. Falsify if NVDA re-accelerates data-center growth into the next print or AMD fails to convert eval interest into cloud design wins.
• Add a tactical long GOOGL vs NVDA pair for 3-6 months. TPU ownership plus vertical stack control should help GOOGL defend inference economics and pricing, while NVDA faces a higher bar on cost-per-token. Cut if GOOGL does not show tangible AI infra monetization or capex discipline deteriorates.
• Treat CRWV and NBIS as watchlist longs, not core longs: buy only on pullbacks after evidence of utilization expansion or margin discipline, ideally via call spreads to cap downside. Backlog alone is not enough; the key risk is that enterprise evaluation intent does not translate into paid, high-occupancy capacity within 1-2 quarters.
• Avoid chasing a direct short NVDA outright ahead of earnings; instead wait for any post-rally failure to hold gains or for channel checks to show pricing pressure. The cleaner falsifier is if NVDA management explicitly offsets with tighter supply, faster attach of software/services, or sustained Blackwell scarcity.