Days before earnings, NVIDIA unveiled an 88-core CPU, a 1.5-ton rack, and an orbital deployment — customers building custom silicon now have to replace three NVIDIAs, not one.
Three identities opened in one week
Ahead of the Wednesday-after-close NVIDIA print, what the market actually got over three days was not a GPU revenue guide — it was an identity re-org. Three distinct product categories were unveiled, and all three carried the same first name: Vera.
First, Vera CPU. 88 cores, 1.2 TB/s memory bandwidth. This is not a GPU-adjacent controller; it is a data-center CPU aimed squarely at the x86/ARM server socket. It is NVIDIA's first commercially shipped CPU as a standalone product line after the Grace era, and the specs put it in the same conversation as flagship server silicon rather than the accelerator sidecar category.
Second, Vera Rubin NVL72. Here the unit of sale is the rack, not the GPU. Taiwanese trade press puts a single rack at over 1.5 tons and roughly $200M sticker, at a scale where deployment requires re-architecting datacenter floor loading and power contracts. Alongside came official token-cost benchmarks for AI agent workloads.
Third, Vera in orbit. SpaceX AI (formerly xAI) selected Vera CPU and a Vera Rubin NVL72 derivative for Grok agent processing and Starmind satellites, with actual orbital deployment locked to Q4 2027. That makes NVIDIA the first semiconductor vendor selling AI compute above the Kármán line.
Why release all three in the same three days
During the same window, most of NVIDIA's largest customers signaled they want out. OpenAI-Broadcom's Jalapeño ASIC published first-party benchmarks claiming 1.9× throughput/watt and 3.6× lower latency versus the GB300. Google committed $200B of capex toward custom silicon expansion. Waymo finished a proprietary robotaxi chip clocking 160 TOPS in real-world use. ARM pivoted from IP vendor to fabricating its own AGI CPU.
Against this, responding with GPU volume and margin alone is a one-day card. NVIDIA's answer is to split the sentence "replace NVIDIA" into three separate replacement projects. To replace the GPU now, a customer also has to replace the CPU, the rack-scale interconnect, and — if that customer is a hyperscaler thinking about frontier deployment — the space-hardened variant. However good the Jalapeño number is, it is one accelerator's story; no custom program yet covers the Vera CPU and NVL72 rack-scale stack in parallel.
The memory side: 12.5× in three years
The cost axis of this identity expansion is DRAM. TrendForce reports Korean DRAM export prices have risen 12.5-fold over three years, outpacing gold; DDR5 16Gb spot sits at $54. Reallocation toward HBM is deepening the commodity DRAM shortage, and the shortage tightens through 2027 rather than easing. A Vera Rubin NVL72 rack materially raises HBM content per system versus the prior generation, which means NVIDIA's capex thesis and Micron/Samsung/SK Hynix's revenue guides now move together on the same fundamental. Micron guided to $50B annual revenue; BMO set a $1,300 target.
But cost is now being pushed to the consumer. Amazon lifted gadget prices citing the AI memory crunch; Apple asked CXMT for a DRAM discount and got Samsung's price instead. Pricing power has moved decisively to the supplier side.
Physical ceiling, orbital exit
A 1.5-ton rack at $200M carries a physical siting constraint. In the same week, Virginia lost 3 GW of datacenter capacity to grid reliability rules. As utility reliability screens tighten, terrestrial deployment schedules become hostage to site and power lead times. That is what makes Starmind interesting beyond the novelty: an orbital deployment opens a TAM outside terrestrial grid regulation. Q4 2027 is not tomorrow, but signaling that a competitor-free market exists off-planet is itself a valuation defense for a stock that just closed seven straight down days.
Positioning
On the call, watch the sequence: (1) GPU revenue and margin, (2) commercial Vera CPU pipeline, (3) NVL72 rack-unit bookings, (4) SpaceX contract color. The market will focus on (1), but the real re-rating trigger is in (2)-(4). Expectations reset lower after the seven-day drawdown, and for the first time this cycle NVIDIA has a direct response to the custom-silicon narrative in the same quarter it faces it.
On memory, this cycle is not near its top. MU delivered a guide raise and target hikes; SK Hynix's 30% New York vs Seoul premium points to global capital in the early innings of re-rating. AVGO and MRVL keep the custom silicon tailwind, but no full-stack alternative covering rack and CPU has appeared on any published roadmap. NVIDIA's bet is that by the time one does, the definition of what they replace will have moved.
Key Sources: - NVIDIA Launches 88-Core Vera CPU with 1.2 TB/s Memory Bandwidth (XFastest, 2026-08-26) - SpaceX Plans AI Satellite Deployment Using NVL72 in Q4 2027 (XenoSpectrum, 2026-08-26) - OpenAI-Broadcom Jalapeño ASIC outpaces Nvidia flagship on power efficiency and latency (Tom's Hardware, 2026-08-25) - South Korea's DRAM Prices Soar 12.5-Fold in 3 Years, Outpacing Gold (TrendForce, 2026-08-26) - NVIDIA Rubin Accelerator Weighs 1.5+ Tons, Reshapes Data Center Deployment (信傳媒, 2026-08-26) - Virginia Loses 3 GW Data Centers to Grid Reliability Concerns (UtilityDive, 2026-08-25) - plus 54 more
If this analysis was helpful · ☕ Support Us · ✈️ Telegram