Mercatus AI

The AI economy runs on tokens. Mercatus discovers their price.

[Mercatus AI @mercatus_ai]

GPU rental prices this week:

B300 $8.89/hr
B200 $6.42/hr
MI300X $4.99/hr
H200 $4.54/hr
H100 $3.85/hr

Datacenter GPUs cost more than they did in July. B300 +5.7%, H100 +4.6% since July 6. Consumer cards slipped.

[Mercatus AI @mercatus_ai]

@nicbstme Hi

[Mercatus AI @mercatus_ai]

@jessiedong_ @lihanc02 Hi I think sector 4 is us

[Mercatus AI @mercatus_ai]

Same open weights, many hosts. Cheapest vs most expensive, per M input tokens:

DeepSeek V4 Flash: 6.5x
GLM-5.2: 3.5x
DeepSeek V4 Pro: 2.9x
Qwen3.8 2.4T: 1.25x

Open weights built the supply side. Nothing yet gives it one price.

[Mercatus AI @mercatus_ai]

Every DeepSeek model, every host, one page.

After the Aug 16 increase, 16 of 17 third-party hosts undercut DeepSeek's own API on V4 Flash. On V4 Pro it depends on the hour.

Official rates, host tables, workload math:
https://t.co/Ei5BrM6lne

[Mercatus AI @mercatus_ai]

Oil, power and freight all have a spot price, a forward curve and a hedge.

LLM tokens have a pricing page.

Same supply structure (many sellers, one product). None of the market.

A posted price you can't lock isn't a cost. It's a forecast.

[Mercatus AI @mercatus_ai]

The harness routes each request to the right model and we at @mercatus_ai gives the inference itself a market.

We’re building the forward market for inference: physically settled token contracts that let providers sell capacity before it exists, and let buyers lock supply before they need it.

[Mercatus AI @mercatus_ai]

DeepSeek, Aug 16: some token prices up 12x, with 10 days' notice.

Same workload, annual cost:
Before: $1,905
After: $3,132 to $6,263
Locked 12mo forward at +5%: ~$2,000

Posted prices move by announcement. Locked prices don't.

https://t.co/quGEeVC6gN

[Mercatus AI @mercatus_ai]

Same week:

Stripe agrees to pay $7B + for OpenRouter (the layer that prices tokens today)

The CFTC opens comment on compute derivatives (the layer that prices them next year)

One exists. One doesn't. That gap is the whole story.

https://t.co/eUNfxMhgFU

[Mercatus AI @mercatus_ai]

US AI inference fell 45.9% in 30 days. It still costs 5x China's.

US TPI: $5.37 per 1M tokens
Coding TPI: $2.63
CN TPI: $1.07

Blended from real usage, not list prices. Rebalanced daily.

https://t.co/FtHxmd92zH

[Mercatus AI @mercatus_ai]

Check token index and compute index today on https://t.co/cJSeNagsbh

[0xSoren @4lili_lili4]

@sui414 https://t.co/oek5dzDPxH

@mercatus_ai's token index is live today and open to everyone. It is still nascent, but we are improving it every day.

[0xSoren @4lili_lili4]

@sui414 https://t.co/oek5dzDPxH

@mercatus_ai's token index is live today and open to everyone. It is still nascent, but we are improving it every day.

[Mercatus AI @mercatus_ai]

DeepSeek just put time-of-day pricing on AI inference.

Aug 16: V4 Pro moves to peak/off-peak billing.

Input: $0.435 → $0.66 off-peak, $1.32 peak
Output: $0.87 → $1.98 / $3.96

Power grids have billed this way for decades.

https://t.co/nh58BGDuGv

[Mercatus AI @mercatus_ai]

Oil until the 1970s. Electricity until the 1990s.

Posted prices, then a spot market, then forward contracts, then a public price curve. Every commodity follows the same path.

AI tokens entered the spot stage in 2026.

[Mercatus AI @mercatus_ai]

US vs Chinese AI inference, per blended 1M tokens:

30 days ago: ~$8.30 vs ~$1.16 (7x gap)
Today: $4.56 vs $1.12 (4x gap)

US prices collapsed 45% in a month. Chinese prices didn't move. Coding inference got MORE expensive (+4.2%).

https://t.co/sbw9MlaEw1

[Mercatus AI @mercatus_ai]

GLM-5.2 trades between $0.07 and $2.10 per million input tokens across 20 providers.

Same weights, 30x apart. The cheapest price fell 95% in 90 days. The official list never moved.

The spread is infrastructure economics, not the product.

[Mercatus AI @mercatus_ai]

GLM-5.2 per 1M input tokens:

Official API: $1.40, unchanged since June
Cheapest host in May: $1.40
Cheapest host today: $0.07

A 95% collapse in 90 days, a 30x spread across 20 hosts. And now the first model with a traded forward price.

https://t.co/mpI0vSWNxs

[Mercatus AI @mercatus_ai]

DeepSeek V3.2, 10M input + 1M output tokens/day, 60% prompt reuse:

Cheapest listed host ($0.20/M in): $2.32/day
"Expensive" official API ($0.28/M in): $1.71/day

The list-price comparison picks the wrong provider by 26%. Cache terms beat sticker price.

https://t.co/u2I8YibN7J

[Mercatus AI @mercatus_ai]

The same open-weight model sells at seven different prices right now.

DeepSeek V3.2: $0.20 to $0.30+ per million input tokens depending on the provider, and $0.18 to $0.57 across the year. The spread is infrastructure economics, not the product.

[Mercatus AI @mercatus_ai]

The same open-weight model can cost 3x more depending on who serves it.

DeepSeek V3.2 runs $0.18 to $0.57 per million input tokens across providers. Same weights, different economics: utilization, precision, caching, margin strategy.

https://t.co/sqLzEyUwqZ