Nvidia’s Two-Speed Offensive: A CPU Push Meets a $150 Billion Taiwan Bet

Nvidia enters the data-center CPU market with Vera, while committing $100-$150B yearly to Taiwan manufacturing and demanding a 20x boost in optical lasers by 2030.

The artificial intelligence hardware race is no longer just about who builds the fastest accelerator. Nvidia is now waging a parallel war on two fronts: carving out a new revenue stream with its first in-house data-center CPU while simultaneously pouring tens of billions of dollars a year into the physical infrastructure needed to keep its expanding portfolio fed. The result is a strategy that ties product ambition directly to manufacturing muscle.

A Network Bottleneck Demands 20x the Lasers

One of the most acute headaches for large-scale AI clusters is shuttling data between thousands of graphics processors. Nvidia is betting on co-packaged optics — optical links placed right next to the chips to deliver higher speeds at lower power. According to a research note from Rosenblatt Securities, the company has told its suppliers it needs a 20-fold increase in capacity for indium-phosphide lasers by 2030. The suppliers themselves view a twelvefold jump as more realistic. The gap underscores how optical networking components have become a choke point for the next wave of AI deployment.

Taiwan at the Center of a $100–$150 Billion Annual Spend

On the hardware production side, the commitment is staggering. CEO Jensen Huang disclosed at the TD Cowen conference on May 28 that Nvidia plans to spend between $100 billion and $150 billion each year with manufacturing partners in Taiwan. Those funds are earmarked for the Blackwell Ultra architecture and the upcoming Vera-Rubin platform, whose first systems are still on track to ship in 2026. The scale of the investment signals that Nvidia intends to own not just the chips but the entire supply chain for its next-generation products.

Vera Exits the Lab

The Vera CPU — first announced in March at GTC in San Jose — has already moved from prototype to production. Huang described it as a multi-billion-dollar business line in its own right. The initial systems went to Anthropic, OpenAI, Oracle Cloud Infrastructure and SpaceXAI, with Nvidia’s Ian Buck personally handing over the hardware. After the AI labs came Oracle’s Santa Clara facility.

Technically, Vera packs 88 custom Olympus cores and a memory bandwidth of 1.2 TB/s. Nvidia is positioning it for agentic AI workloads: orchestration, tool calls, reinforcement learning and managing long context windows. It is also the first data-center CPU to use LPDDR5 memory, which the company argues will deliver significantly more performance per watt.

Oracle, CoreWeave and the Cloud Rollout

Oracle Cloud Infrastructure plans to deploy hundreds of thousands of Vera CPUs starting in 2026, making OCI the first hyperscaler to adopt the chip at scale. CoreWeave has been confirmed as the initial cloud customer for standalone Vera access. Broader availability through AWS, Google Cloud, Azure, Lambda and Nebius is expected in the second half of the year.

Beyond its standalone role, Vera will serve as the host processor inside the Vera-Rubin NVL72 systems, connecting to Rubin GPUs via NVLink-C2C. The move is part of a broader strategy to take control of more sections of the AI data center rather than just delivering accelerators.

Analyst Math: A $24.6 Billion CPU Market by 2028

Wolfe Research estimates that agentic AI will expand the CPU market by roughly 30 percent by 2028. The logic is simple: modern AI systems need more traditional processor power to coordinate tasks, manage memory and steer complex workflows alongside GPUs.

For Nvidia, Wolfe expects more than 4 million CPUs shipped this year, of which about 1.3 million will be Vera processors for agentic AI — most of those in the fourth quarter. Revenue from agentic CPUs is projected to leap from $6.6 billion in 2026 to $24.6 billion by 2028.

Yet the context matters. Compared with Nvidia’s dominant accelerator business, the CPU division will remain smaller for now. Wolfe sees AMD — which is far more leveraged to server CPUs relative to its own size — as particularly well placed. Its server CPU revenue could climb from roughly $17 billion in 2026 to $44 billion by 2028. The wildcard is not chip performance but TSMC’s fabrication capacity, which could determine market share more decisively than any datasheet metric.

Margins, Guidance and the Stock’s Cool-Down

Operationally, Nvidia continues to run a non-GAAP gross margin near 75 percent. Some analysts at the Cowen conference flagged rising memory costs as a potential headwind, though the high margins from Nvidia’s software ecosystem and the new CPU business are seen as offsets.

For the second quarter of 2026, the company guided for revenue of roughly $91 billion — above the prior market consensus of $87.2 billion — representing year-over-year growth of 95 percent. The first quarter had already delivered $81.6 billion.

The stock, meanwhile, has cooled after a strong run. Trading in Frankfurt at €184.68, it sits 8.14 percent below its 52-week high, though it has still gained 14.64 percent year-to-date. The relative strength index of 36.4 suggests the shares are neither overheated nor deeply oversold. Over twelve months the gain stands at 51.38 percent, and the distance to the 200-day moving line is 14.98 percent. Among the 43 analysts covering Nvidia, the average view remains “Buy” with a median 12-month price target of $304.59.

The Next Catalyst: GTC Taipei on June 1

The upcoming milestone for investors is the GTC keynote in Taipei on June 1, where Huang is expected to reveal more details on the Vera-Rubin architecture and potentially introduce new products aimed at physical AI and robotics. For Nvidia, the second half of the year will be the true test: if Vera can move beyond pilot customers into scalable cloud demand, the CPU campaign will become another credible leg of the platform story — one that already includes a $150 billion annual bet on Taiwan’s factories.