Wall Street comments on GTC: In NVIDIA's definition, computing power is revenue, and tokens are the new commodities.

```

Nvidia's annual GTC conference delivered a core signal: the business logic of AI computing power is undergoing a fundamental reconstruction—tokens have become the new commodity, and computing power equals revenue.

At this GTC, Nvidia management sharply raised its visibility on data center sales from $500 billion (covering through 2026) to over $1 trillion (cumulative from 2025 to 2027), and made it clear that sales of the standalone Vera CPU and LPX rack solutions will be counted as additional revenue on top of this. Wall Street regards this conference as a strong endorsement of the continuity of Nvidia's AI cycle.

According to Chasewind Trading Desk, JPMorgan's latest report points out that this figure means there is at least a $50 to $70 billion upside versus Wall Street's current consensus expectations for data center revenue in 2026-2027.

BofA Securities directly quoted Nvidia management—"Tokens are the new commodity, and computing power equals revenue"—and pointed out that the Blackwell system achieves up to 35x lower cost per token compared to the previous generation Hopper, and the upcoming Rubin series is expected to further lower this by 2x to 35x, with the specific range depending on workload type and architecture configuration.

Within Nvidia's narrative framework, this ever-shrinking token cost curve is the fundamental driver for the scaling up of demand.

Doubling Visibility of Demand, Dual Drivers from Hyperscale Customers and Enterprise Market

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

The demand structure is becoming diversified: about 60% comes from hyperscale cloud vendors (internal AI consumption is migrating from recommendation/search workloads to large language models), and the remaining ~40% is distributed among CUDA cloud-native AI enterprises, Nvidia cloud partners, sovereign AI, and industrial/enterprise customers.

BofA points out that this new $1 trillion outlook essentially matches the previous Wall Street expectation of about $970 billion in data center revenue for this three-year period, validating logic much the same as the previous validation of the $500 billion old estimate and the $450 billion expected.

Notably, Nvidia management devoted considerable time in this conference to explaining the acceleration of traditional enterprise workloads as a new vector of demand.

Nvidia announced partnerships with IBM (accelerating WatsonX), Google Cloud (BigQuery acceleration, Snap achieving about 76% cost savings), Dell (AI data platform), and launched the cuDF and cuVS CUDA-X foundational libraries.

JPMorgan believes this direction is "seriously underestimated by the market"—because Moore's Law is coming to an end, domain-specific acceleration is the only viable path, and this will expand Nvidia's addressable market beyond just AI training/inference cycles.

Groq LPU Integration: The Most Important New Architecture Release

JPMorgan rated the integration of the Groq 3 LPU with Vera Rubin as the "most important new architecture release" at this GTC.

This decoupled inference 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): prefill is done on Vera Rubin, attention decoding is also run on Rubin, while feedforward networks/token generation is offloaded to the Groq LPU.

The LPX rack integrates 256 LPUs, providing 128GB aggregated SRAM, 40PB/s memory bandwidth, and 315 PFLOPS inference computing power, with a market launch expected in Q3 2026.

Nvidia management stated that for workloads requiring ultra-high token speeds (code generation, engineering computing, long-context inference), about 25% of data center power will be allocated to LPX, and the remaining 75% to pure Vera Rubin NVL72 configurations.

BofA data shows that after pairing the Rubin system with the SRAM LPX rack, efficiency for high-end, low-latency workloads can improve by up to 35x compared to the previous generation. JPMorgan 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), enabling Nvidia to effectively compete in the high-end inference market where ASIC vendors have traditionally held an edge.

Copper Cable and CPO Advancing in Parallel, No Single Bet on Interconnect Approach

Nvidia management directly addressed the debate between copper cable and co-packaged optics (CPO) at the conference, confirming that both routes will be pursued in parallel.

Currently, in the Vera Rubin generation, Oberon racks use copper cable scaling up to NVL72, and optical expansion up to NVL576; the Spectrum-6 SPX co-packaged optical Ethernet switch is already in mass production, jointly developed by Nvidia and TSMC. Management claims its power efficiency is five times higher than traditional pluggable transceivers, with reliability improved tenfold.

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

BofA emphasized that for CPO expansion/horizontal switches, adoption is totally optional for customers—clients can continue using copper up to when they deem appropriate. JPMorgan believes this dual-path confirmation is consistent with their previous predictions, expecting copper will continue to dominate NVL72/NVL144 configurations through at least 2027, with CPO gradually increasing its share in horizontal scaling and NVL576+ configurations.

Vera CPU: A New Billions-Scale Revenue Source for Agent AI

Nvidia management stated at the conference that the Vera CPU standalone business "is already set to become a multi-billion dollar business," and BofA notes that this has not yet been factored into current consensus market expectations—representing incremental contributions.

The Vera CPU is equipped with 88 proprietary Olympus ARM cores, an LPDDR5X memory subsystem with 1.2TB/s bandwidth (only half the power consumption of traditional server CPUs), and connects to the GPU via NVLink-C2C at 1.8TB/s (7x the speed of PCIe Gen 6). The Vera CPU rack integrates 256 liquid-cooled CPUs, supporting over 22,500 concurrent CPU environments.

Management emphasized that the CPU is becoming the bottleneck for scaling agent AI—reinforcement learning and agent workflows require large numbers of CPU environments to test and validate the GPU model outputs. Meta is already deploying the previous generation Grace CPU at scale, with Vera set to take over in 2027.

JPMorgan characterizes this CPU revenue stream as high-gross-margin, repeatable (being deployed along with GPU racks in AI factories), and structurally tied to the agent AI adoption curve that Nvidia is actively catalyzing.

Product Roadmap Extended to 2028, Annual Architecture Cadence Strengthened

Nvidia reaffirmed its annual platform launch cadence: Blackwell (2024) → Blackwell Ultra (2025) → Rubin (2026) → Rubin Ultra (2027) → Feynman (2028).

Rubin Ultra will adopt a four-chip GPU setup, equipped with 1TB HBM4e, and introduce the new LP35 LPU chip (first time featuring NVFP4 performance). The Kyber rack supports up to 144 GPUs per NVLink domain (seventh generation NVLink, 3.6Tb/s per GPU, NVL576 aggregate bandwidth of 1.5Pb/s).

Details about Feynman exceed market expectations:

The new GPU uses TSMC's A16 (1.6nm) process, introduces chip stacking and custom HBM; the new CPU is called Rosa (named after Rosalind Franklin), designed specifically for orchestrating agent workloads across GPU, LPU, storage, and networking; the new LPU is called LP40, co-developed by Nvidia's internal Groq team; in addition, there is the BlueField-5 DPU, ConnectX-10 super NIC, NVLink 8, and Spectrum-7 (204T, CPO).

JPMorgan believes that Nvidia's vertically integrated platform (now spanning seven chips, five rack systems, and supporting software stacks) will be difficult to replicate, and that accelerated inference demand, the expansion of addressable markets driven by accelerating traditional workloads, and continued broadening of the customer base together sustain a more durable AI capital expenditure cycle than the market currently expects.

~~~~~~~~~~~~~~~~~~~~~~~~

The above content is from Chasewind Trading Desk.

For more detailed interpretation, including real-time analysis and frontline research, please join [Chasewind Trading Desk Annual Membership]

Risk Warning and DisclaimerThe market is risky, investment needs caution. This article does not constitute personal investment advice, nor does it consider the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, viewpoints, or conclusions in this article are suitable for their particular situation. Investing based on this is at your own risk. ```