Feb 17 – 23, 2026
H100 resale prices crater 85% as Blackwell takes over. The SaaSpocalypse wipes $2T from software stocks. NVIDIA's pivotal Q4 earnings loom. Three labs ship four new models in a single month. Here's everything that moved the AI compute market this week.
The Yotta GPU Index declined 3.8% this week, primarily driven by continued downward pressure on Hopper-generation pricing as Blackwell availability expands. Conversely, the Yotta Model Index rose 5.2% as new model releases (Claude Sonnet 4.6, GPT-5.3-Codex) temporarily increased premium-tier token demand. The composite Yotta Compute Index edged down 1.4%, reflecting the net deflationary trend in AI compute costs.
Hopper-generation GPUs (H100, H200) continued their downward trajectory this week as Blackwell supply ramps up. The H200 saw the steepest decline at -6.2%, while newer Blackwell GPUs (B200) surged 8.4% on constrained supply and strong demand. Consumer-grade Ada Lovelace GPUs (L40S, RTX 4090) posted modest gains as inference workloads shift to cost-efficient hardware.
| GPU | Arch | Avg $/hr | 7d Change |
|---|---|---|---|
| H100 80GB | Hopper | $2.85 | -4.7% |
| H200 141GB | Hopper | $4.10 | -6.2% |
| A100 80GB | Ampere | $1.42 | -2.1% |
| A100 40GB | Ampere | $1.05 | -1.9% |
| L40S 48GB | Ada Lovelace | $0.89 | +1.1% |
| RTX 4090 24GB | Ada Lovelace | $0.44 | +2.3% |
| B200 192GB | Blackwell | $5.80 | +8.4% |
| RTX 5090 32GB | Blackwell | $0.72 | +4.3% |
The H100 resale market has entered a structural decline phase. With Blackwell B200/B300 offering 11–98x better inference economics, we expect Hopper-generation spot prices to stabilize around $2.00–$2.50/hr by Q2 2026. Meanwhile, B200 supply constraints will keep next-gen pricing elevated through mid-year. Watch NVIDIA's Q4 earnings (Feb 25) for updated Blackwell production guidance.
February 2026 has been the most active month for model releases in recent memory. Anthropic's Claude Sonnet 4.6 and OpenAI's GPT-5.3-Codex both launched this month, while DeepSeek accelerates its R2 model timeline. Token pricing remains stable for established models, with competitive pressure driving down costs for older-generation flagship models like Claude Opus 4.6 (-5.0%) and Llama 3.1 405B (-12.5%).
| Model | Provider | Input $/M | Output $/M | 7d Change |
|---|---|---|---|---|
| GPT-4o | OpenAI | $2.50 | $10.00 | 0.0% |
| GPT-4o Mini | OpenAI | $0.15 | $0.60 | 0.0% |
| GPT-5.3-CodexNew | OpenAI | $5.00 | $20.00 | -8.3% |
| Claude Sonnet 4.6New | Anthropic | $3.00 | $15.00 | 0.0% |
| Claude Opus 4.6 | Anthropic | $5.00 | $25.00 | -5.0% |
| DeepSeek V3 | DeepSeek | $0.27 | $1.10 | 0.0% |
| Gemini 2.0 Flash | $0.07 | $0.30 | 0.0% | |
| Llama 3.1 405B | Meta | $0.80 | $0.80 | -12.5% |
NVIDIA's H100 GPU, once dubbed "digital gold," has seen its resale price collapse from $40,000 to approximately $6,000 per unit. The primary driver is the Blackwell architecture (B300/B200), which offers up to 98x better inference performance. Tech giants including Microsoft, Google, and Meta have extended GPU depreciation periods to 5–6 years to mask the real asset devaluation.
Software market caps lost approximately $2 trillion as investors reckon with the disruptive potential of LLMs replacing current SaaS offerings. Deutsche Bank noted that markets had been "implicitly pricing in a world where almost every tech company would come out a winner." Legal, IT, consulting, and logistics sectors were particularly impacted.
NVIDIA reports Q4 fiscal 2026 results on February 25. Wall Street expects revenue of $65 billion and EPS of $1.52 (up 71% YoY). The Blackwell Ultra architecture delivers up to 50x better performance than Hopper. The upcoming Vera Rubin GPU is scheduled for later in 2026, featuring NVIDIA's first custom CPU design.
Anthropic launched Claude Sonnet 4.6 (Feb 18) with major upgrades across coding, computer use, and long-reasoning. OpenAI released GPT-5.3-Codex targeting developer workflows. Meanwhile, DeepSeek is accelerating its R2 model (142B parameters) which has already outperformed Llama-3.1-405B and GPT-4o in benchmarks — with only 1.2% additional computational overhead.
The five largest American tech companies are set to spend $700 billion in capital expenditure in 2026, primarily on AI data center infrastructure. OpenAI alone projects $600 billion in compute spending through 2030. QumulusAI deployed 1,144 NVIDIA Blackwell GPUs, with B300 and RTX Pro 6000 rollouts planned throughout 2026.
AWS Inferentia2 chips can now run large language models at 70% lower cost than traditional GPU instances. The inference market is becoming the next AI chip battleground, with growing enterprise demand creating mainstream opportunities for GPU alternatives. This shift is pressuring traditional cloud GPU pricing downward.
Join 12,000+ AI compute professionals who receive our market intelligence every Sunday.
No spam. Unsubscribe anytime.
Disclaimer: This report is for informational purposes only and does not constitute financial advice. GPU pricing data is aggregated from multiple cloud providers and may not reflect real-time spot prices. Model token pricing is based on published API rates. Prediction market data reflects community sentiment and is not indicative of actual outcomes. Past performance is not indicative of future results.