Few product cycles have generated as much anticipation—or revenue—as Nvidia's Blackwell architecture. The B200 GPU, now shipping to hyperscalers worldwide, is the centerpiece of what Jensen Huang calls “the first $1 trillion data center buildout in history.”

>

The Blackwell B200: Specs, Pricing, and Production

The B200 packs 208 billion transistors on TSMC's 4NP process, delivering 4.5 PFLOPS FP8 Tensor Core performance (2.5x vs H100), 192 GB of HBM3e memory at 8 TB/s, and NVLink 5.0 at 1.8 TB/s per GPU. Enterprises are paying $30,000–$50,000 per GPU, with DGX B200 systems north of $300,000. Demand outstrips supply: lead times run 36–52 weeks. Current estimates point to roughly 1.5–2 million B200 units shipping in calendar 2026, implying over $50 billion in Blackwell GPU revenue alone. TSMC's CoWoS-L packaging remains the primary bottleneck, though additional capacity from Amkor's new Arizona plant is expected in H2 2026.

Hyperscaler Demand: Who's Buying and How Much?

The ramp is driven overwhelmingly by the four major US hyperscalers, all accelerating CapEx for 2026:

  • Microsoft — deploying B200s across Azure for Copilot, Bing, and OpenAI inference, with AI infrastructure spend tracking toward $60B+
  • Google — adopting Blackwell alongside internal TPU v6, with 2026 CapEx guided at $55B+
  • Amazon — offering B200 through EC2 P6, SageMaker, and Bedrock, with CapEx expected to exceed $65B
  • Meta — committed to Blackwell for Llama 4 training and recommendation systems, with CapEx guided to $40B+

Beyond the hyperscalers, enterprise adoption is accelerating across financial services, healthcare, and energy. The “sovereign AI” wave—Japan, India, Saudi Arabia, and the UAE building national AI supercomputers—adds another demand layer.

▶ CNBC
Nvidia Vera Rubin and Blackwell analysis

Competition: AMD, Cerebras, and Custom ASICs

AMD's /ticker/AMD MI350, based on CDNA 4 and shipping late 2025, narrows the raw performance gap with up to 2.3x FP8 improvement over MI300X and 288 GB of HBM3e. AMD guided $15–18B in data center GPU revenue for 2026, but supply remains a major headwind: TSMC CoWoS allocation for AMD is roughly one-sixth of Nvidia's. ROCm software has made meaningful strides but still lags CUDA.

Cerebras has carved an inference niche with wafer-scale CS-3, claiming up to 10x lower latency vs H100 for LLMs, but remains a small player ($800M–1.2B estimated 2026 revenue). The more credible long-term threat is custom ASICs: Google's TPU v6, Amazon's Trainium 3, and Microsoft's Maia 100 are designed to reduce Nvidia dependence for specific internal workloads. We estimate custom ASICs will reach 15–20% of AI accelerator spending by 2028, up from ~8% today.

NVDA Valuation: The ‘70% CAGR’ Debate

Nvidia's data center revenue has grown from $15B in FY2024 to an estimated $78B in FY2026, on track for $105–120B in FY2027. Gross margins sit at 78–80% with operating margins near 62%. At ~$1,150, /ticker/NVDA trades at ~32x forward P/E with a PEG ratio of ~0.9x — defensible but not cheap.

The bull case argues AI compute is a structural multi-decade shift. The bear case points to inference efficiency gains, model commoditization, and capex ROI concerns that could trigger a correction. Our view: the bull case holds for 12–18 months, but the “70% CAGR” will decelerate to 25–35% growth from FY2028 onward as the base grows and competitive alternatives gain share.

“Nvidia’s competitive moat — spanning hardware, networking, and the CUDA ecosystem (4M+ developers) — will limit share losses to non-core workloads where hyperscalers can achieve meaningful savings through vertical integration.”

Beyond Blackwell, the Rubin architecture (2027, TSMC N3, up to 300B transistors, HBM4) and Spectrum-X networking (a $10B+ opportunity) extend the runway. We remain bullish on NVDA with a clear stop-loss framework — the franchise remains the defining bet of the AI era.

Disclosure: The Signal may hold positions in securities mentioned. This article is for informational purposes only and does not constitute investment advice.