The same chip.
$3.50 to $14.24.

As of July 2026, on-demand NVIDIA B200 capacity trades at a median of roughly $6.12–$6.25 per GPU-hour across independent multi-provider price trackers. The published on-demand range runs from about $3.50/GPU-hr at aggregators such as Vultr and Hyperbolic up to $14.24/GPU-hr at AWS, whose p6-b200.48xlarge instance lists at $113.93/hour for eight B200 GPUs. Mid-market specialist clouds cluster between $5.89 and $8.60: RunPod at $5.89, Lambda at $6.69, Nebius at $7.15, and CoreWeave at $8.60. SmartTec's published reserved rate for B200 capacity at its Mead, Oklahoma site is $4.50/GPU-hr, with power-on targeted for Q4 2026.

Because the price reflects the provider's cost stack, not the silicon. A hyperscaler carries metro land, retail power, global redundancy, enterprise sales and support organizations, and egress economics — AWS lists B200 at $14.24/GPU-hr because that structure requires it. A specialist neocloud strips out the general-purpose cloud and lands between $5.89 and $8.60. A marketplace resells other people's idle racks and can go below $4. SmartTec's cost stack is unusually low for a different reason: it owns its 30-acre Mead site and 3 MVA transformer outright, buys Oklahoma grid power at roughly $0.08/kWh against $0.12–0.25/kWh for leased colocation, and sources its z1power LFP battery storage at manufacturer cost from a company its founders own.

No, and the difference is the product rather than the chip. Sub-$4 B200 listings are almost always shared-infrastructure or marketplace capacity: preemptible in practice, variable in performance, without a dedicated high-speed fabric, and sold by hosts whose reliability varies widely. Rates in the $5.50–$8.60 band typically buy dedicated hardware, a private InfiniBand fabric, a support relationship, and a contract. SmartTec sells the second kind — its $4.50/GPU-hr reserved rate covers dedicated B200s on a private fabric with a named engineer, which is why the honest comparison for that price is Lambda or CoreWeave reserved capacity, not a spot marketplace listing.

SmartTec's $4.50/GPU-hr reserved rate sits roughly 28% under the July 2026 market median of $6.25 because most of its cost stack is owned rather than rented. The company owns 30 acres in Mead, Oklahoma and a 3 MVA on-site transformer, which removes both land rent and the multi-year interconnection queue that gates most new capacity — Phase 1A's 114 kW load uses about 4% of that transformer. Power comes from the Oklahoma grid at approximately $0.08/kWh, battery storage comes from founder-owned z1power at manufacturer cost, and 100 Gbps symmetrical fiber is locked on a signed 60-month Dobson quote at $8,075/month. A competitor must buy at retail what SmartTec already holds at cost.

Most independent forecasts point down as supply arrives — some project B200 on-demand rates softening toward $2.50–$3.00/GPU-hr at scale by late 2026 as backordered units ship, and the historical precedent is that H100 rates fell roughly 35% over two years. SmartTec models this explicitly rather than assuming rates hold: its financial base case applies an annual rate decline, and Phase 1 revenue is contracted at a fixed rate for year one so that decay only affects renewals. Entry price on the hardware matters more than the decay slope, which is why SmartTec's model is built on refurbished HGX B200 systems rather than retail.