Wall Street Commentary GTC: In NVIDIA's definition, computing power equals revenue, and Token is the new commodity

NVIDIA's annual GTC conference released a core signal: the commercial logic of AI computing power is undergoing a fundamental reconstruction—Token has become the new commodity, and computing power equals revenue.

At this year's GTC, NVIDIA's management significantly raised the visibility of data center sales from the previous $500 billion (covering until 2026) to over $1 trillion (covering a cumulative period from 2025 to 2027), and clearly stated that the sales of the independent Vera CPU and LPX rack solutions will be counted additionally. Wall Street views this conference as a strong endorsement of the sustainability of NVIDIA's AI cycle.

According to the Wind Trading Desk, JP Morgan's latest report indicates that this figure implies at least $50 billion to $70 billion of upside potential compared to Wall Street's current consensus expectation for data center revenue in 2026 to 2027.

The Bank of America report directly quoted NVIDIA's management statement—"Token is the new commodity, computing power equals revenue"—and pointed out that the Blackwell system has achieved a cost reduction of up to 35 times per Token compared to the previous generation Hopper, and the upcoming Rubin series is expected to further reduce this by 2 to 35 times, depending on the workload type and architecture configuration.

Within NVIDIA's narrative framework, this continuously compressing Token cost curve is the fundamental driving force behind the scaling of demand.

Demand visibility doubles, driven by both hyperscale customers and enterprise markets

NVIDIA's management disclosed that high-confidence purchase orders for the Blackwell and Vera Rubin systems have exceeded $1 trillion, doubling from the $500 billion announced at the GTC DC conference in October 2025. Management also indicated that additional orders and backlogs for 2027 are expected to continue accumulating over the next 6 to 9 months.

The demand structure is diversifying: about 60% comes from hyperscale cloud providers (internal AI consumption is shifting from recommendation/search workloads to large language models), while the remaining approximately 40% is distributed among CUDA cloud-native AI companies, NVIDIA cloud partners, sovereign AI, and industrial/enterprise customers.

The Bank of America noted that this new $1 trillion outlook is basically in line with Wall Street's previous expectation of about $970 billion for the three-year data center revenue, validating the logic in the same way as the old $500 billion outlook validated the expectation of about $450 billion in October 2025.

It is noteworthy that NVIDIA's management elaborated on the acceleration of traditional enterprise workload as a demand vector during this conference.

NVIDIA announced collaborations with IBM (accelerating WatsonX), Google Cloud (BigQuery acceleration, Snap achieving approximately 76% cost savings), Dell (AI data platform), and launched two major CUDA-X foundational libraries, cuDF and cuVS.

JP Morgan believes this direction is "severely underestimated" by the market—the logic being that Moore's Law has become ineffective, and domain-specific acceleration is the only viable alternative path, which will expand NVIDIA's addressable market beyond AI training/inference cycles

Groq LPU Integration: The Most Important New Product Release at the Architectural Level

JP Morgan rated the integration of Groq 3 LPU with Vera Rubin as the "most important new product release at the architectural level" at this year's GTC.

This decoupled reasoning architecture pairs the Rubin GPU (high throughput, 288GB HBM4, 22TB/s bandwidth, 50 PFLOPS NVFP4) with the Groq LPU (low latency decoding, 500MB on-chip SRAM, 150TB/s SRAM bandwidth, 1.2 PFLOPS FP8): pre-filling is completed on Vera Rubin, the attention decoding part also runs on Rubin, while the feedforward network/Token generation is offloaded to Groq LPU.

The LPX rack integrates 256 LPUs, providing 128GB aggregated SRAM, 40PB/s memory bandwidth, and 315 PFLOPS inference computing power, expected to be launched in the third quarter of 2026.

NVIDIA's management stated that for workloads requiring ultra-high Token speeds (code generation, engineering calculations, long context reasoning), about 25% of data center power consumption will be allocated to LPX, with the remaining 75% being pure Vera Rubin NVL72 configuration.

Bank of America data shows that the efficiency of high-end low-latency workloads with the Rubin system paired with SRAM LPX rack can improve by 35 times compared to the previous generation. JP Morgan pointed out that this architecture directly addresses the fundamental contradiction that a single processor cannot simultaneously optimize throughput (limited by FLOPS) and latency (limited by bandwidth), allowing NVIDIA to effectively compete in the high-end inference market, where ASIC manufacturers previously had traditional advantages.

Copper Cables and CPO Progressing in Parallel, No Single Bet on Interconnection Routes

NVIDIA's management directly addressed the competition between copper cables and Co-Packaged Optics (CPO) at the conference, confirming that both routes will be advanced simultaneously.

In the current Vera Rubin generation, the Oberon rack uses copper cables to extend to NVL72 and optical to NVL576; the Spectrum-6 SPX co-packaged optical Ethernet switch has been mass-produced, jointly developed by NVIDIA and TSMC, with management stating that its optical power efficiency is 5 times better than traditional pluggable transceivers, and resilience is improved by 10 times.

For Rubin Ultra (second half of 2027), the Kyber rack will use copper cable NVLink expansion (up to 144 GPUs), while also providing a CPO-based NVLink switching solution as an alternative. Feynman (2028) will explicitly support both copper cable and CPO expansion, equipped with Spectrum-7 (204T, CPO) for horizontal expansion.

Bank of America emphasized that the adoption of CPO expansion/horizontal expansion switches is entirely optional for customers, who can continue to use copper cables until they deem it appropriate. JP Morgan believes that this dual-path confirmation aligns with its previous predictions, expecting copper cable expansion to continue to dominate NVL72/NVL144 configurations at least until 2027, while CPO will gradually expand its share in horizontal expansion and NVL576+ configurations

Vera CPU: A New Independent Billion-Dollar Revenue Source for Intelligent AI

NVIDIA's management clearly stated at the conference that the independent business of Vera CPU "has been confirmed to become a billion-dollar scale business." Bank of America Securities pointed out that this revenue stream has not yet been covered by current market consensus expectations and represents incremental contribution.

The Vera CPU is equipped with 88 self-developed Olympus ARM cores, and the LPDDR5X memory subsystem provides 1.2TB/s bandwidth (with power consumption only half that of traditional server CPUs), and interconnects with GPUs at a rate of 1.8TB/s via NVLink-C2C (equivalent to 7 times PCIe Gen 6). The Vera CPU rack integrates 256 liquid-cooled CPUs, supporting over 22,500 concurrent CPU environments.

Management emphasized that CPUs are becoming a bottleneck for the expansion of intelligent AI—reinforcement learning and intelligent workflows require a large number of CPU environments to test and validate the output results of GPU models. Meta has already scaled the deployment of the previous generation Grace CPU, and Vera will replace it in 2027.

JP Morgan characterizes this CPU revenue stream as high-margin and repeatable (deployed alongside GPU racks in AI factories), forming a structural binding with the adoption curve of intelligent AI that NVIDIA is actively catalyzing.

Product Roadmap Extended to 2028, Annual Architecture Rhythm Continues to Strengthen

NVIDIA reiterated its annual platform release rhythm: Blackwell (2024) → Blackwell Ultra (2025) → Rubin (2026) → Rubin Ultra (2027) → Feynman (2028).

Rubin Ultra will adopt a 4-chip GPU configuration, equipped with 1TB HBM4e, and introduce the LP35 LPU chip (first introducing NVFP4 computing power). The Kyber rack supports 144 GPUs per NVLink domain (seventh-generation NVLink, 3.6Tb/s per GPU, NVL576 aggregated bandwidth of 1.5Pb/s).

Details of Feynman’s disclosure exceeded market expectations:

The new GPU uses TSMC's A16 (1.6nm) process, introducing chip stacking and customized HBM; the new CPU is named Rosa (after Rosalind Franklin), designed specifically for orchestrating intelligent workloads across GPUs, LPUs, storage, and networks; the new LPU is named LP40, co-developed by NVIDIA's internal Groq team; it also includes BlueField-5 DPU, ConnectX-10 super network card, NVLink 8, and Spectrum-7 (204T, CPO).

JP Morgan believes that NVIDIA's vertically integrated platform (now spanning seven chips, five rack systems, and supporting software stacks) is difficult to replicate. The acceleration of inference demand, structural expansion of the addressable market brought by traditional workload acceleration, and the continuous broadening of the customer base together support a more sustainable AI capital expenditure cycle than the market currently expects