Nvidia Blackwell GPUs Set to Slash AI Token Costs by 35x

The cost of running AI models is about to drop dramatically as Nvidia's Blackwell GPU systems reach scale in data centers, potentially reshaping pricing across the industry.

According to analysis from SemiAnalysis, Nvidia's top-tier Blackwell system—the GB300 NVL72—generates tokens at roughly one-thirty-fifth the cost of its predecessor, the Hopper HGX 200. The older system cost $4.20 per million tokens, while Blackwell delivers the same output for just 12 cents.

Tokens are the fundamental units AI models use to process information, and they serve as the standard measure for pricing AI services. The dramatic efficiency gains from Blackwell hardware are expected to flood the market with cheaper tokens as installations accelerate through 2026.

Performance leap beyond raw speed

The Blackwell advantage extends beyond simple cost reduction. Each GPU in the new system generates 6,000 tokens per second compared to 90 tokens per second in Hopper—a 65-fold increase in raw throughput.

But energy efficiency tells an even more compelling story. When measured by tokens generated per megawatt of electricity consumed, Blackwell produces 2.8 million tokens per second versus Hopper's 54,000. That 50-times improvement matters increasingly as electricity costs rise due to surging data center demand.

These systems represent genuine supercomputers rather than simple chip upgrades. Their deployment required significant infrastructure changes, including water cooling systems and other complex data center modifications that delayed initial rollouts.

Why it matters

The token cost collapse could fundamentally alter AI economics. Companies that have been rationing AI usage due to expense may suddenly find large-scale deployment viable. Conversely, the abundance of cheap tokens might trigger a new wave of profligate consumption that offsets efficiency gains—a phenomenon some have dubbed "tokenmaxxing." Either way, the competitive landscape for AI model providers will shift as cost advantages from proprietary infrastructure erode.

Price cuts already beginning

Several AI model providers have already begun reducing prices in anticipation of the efficiency gains. OpenAI CEO Sam Altman recently acknowledged that AI costs had become "a huge issue" and promised the company would offer "a lot of ways we can help people get more value for less spend."

Early market signals support the trend. Silicon Data's closely watched token spending index peaked at 2.06 in late May 2026 before falling to 1.75 by June 10. Carmen Li, CEO of Silicon Data, indicated this decline likely reflects dropping token prices across multiple AI models.

An AI infrastructure CEO, speaking anonymously over lunch, predicted that new AI models launching later in 2026 would be "a lot better and more efficient," directly enabling the token price reductions.

As Blackwell systems continue rolling out at scale through the second half of 2026, the massive increase in cheaply-generated tokens will give providers both the capability and competitive pressure to slash prices further.

These details were first reported by Alistair Barr at Business Insider.