NVIDIA’s Blackwell family – the B200, the Blackwell Ultra B300, the rack-scale DGX B300 system, and the 72-GPU NVL72 racks (GB200 NVL72 and the newer GB300 NVL72) – sits at the center of AI infrastructure spending in 2026. Buyers are juggling on-prem capex, cloud GPU-hour rates, and rapidly shifting availability. This breakdown collects the latest verified pricing signals, updated for August 2026, so you can benchmark quotes against the live market.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Blackwell Pricing at a Glance
For readers who want the headline numbers before the sourcing detail, here’s every Blackwell price point in this guide as of August 2026:
- B200 cloud rate: $5.91–$16.11 per GPU-hour across 11 named providers in the latest August 2026 cloud comparison, from Vast.ai’s floor to Google Cloud’s hyperscaler ceiling – or as low as $5.29/hr at Lambda Labs per GridStackHub’s narrower 6+ provider index
- B300 cloud rate: $7.10–$17.80 per GPU-hour across five named on-demand providers, from Modal’s floor to AWS’s ceiling – premium managed DGX B300 stacks run as high as $18.00/hr, or as low as $3.13/hr on a 48-month reservation
- DGX B200 system (8 GPUs): $280,000–$320,000
- DGX B300 system (8 GPUs): $300,000–$350,000
- DGX Station (1 GPU, desktop): $80,000–$125,000
- GB200 NVL72 rack (72 GPUs, estimate): $2 million–$3 million
- GB300 NVL72 rack (72 GPUs, estimate): $3 million–$4 million
The rest of this guide breaks down where each figure comes from, how the tiers compare, and when renting beats buying.
Where These Prices Come From
This guide anchors its figures to named, dated sources rather than generic estimates: NVIDIA’s own DGX B300 product page, live on nvidia.com as of August 2026, for system specs and positioning; iFactory App’s DGX B200-vs-B300 comparison (Q1–April 2026) for system-level pricing, corroborated by IntuitionLabs’ figures for the DGX B300 anchor, IntuitionLabs’ August 2026 DGX Station pricing, and Thunder Compute’s DGX B200 MSRP corroboration; an August 2026 cloud GPU-hour comparison spanning 11 named providers – Vast.ai, Hyperbolic, Hyperstack, Lambda, Runpod, Nebius, Vultr, CoreWeave, Oracle Cloud, AWS, and Google Cloud – for current B200 pricing, a broader read than GridStackHub’s earlier August 2026 index of 6+ specialized providers; and Thunder Compute’s August 2026 pricing table, naming Modal, Runpod, Nebius, Oracle Cloud, and AWS, plus a separate August 2026 index naming Spheron’s on-demand SXM6 listing and premium managed DGX B300 stacks, for current B300 GPU-hour pricing. NVIDIA’s DGX B300 page confirms specs but still does not list a system price or a per-GPU retail price for B200 or B300 units – every per-GPU and system figure in this article remains an OEM quote or early-listing estimate, not a manufacturer list price. Earlier-2026 cloud pricing from Spheron Network, Verda, and GPUAAS – including a B200 floor near $2.71/hr and B300 spot pricing near $2.45/hr – has since been superseded by the tighter, higher August 2026 readings referenced throughout this guide; GPUAAS’s July 2026 wholesale and hyperscaler B300 bands still corroborate the current named-provider figures, so they remain referenced below. A separate August 2026 pricing guide contributes the 48-month reserved B300 rate. The GB300 NVL72 rack price is sourced from broader market estimates rather than these primary named sources and carries lower confidence until a dated, named quote is available. A separate July 2026 pricing guide contributes the lower B200 street-price estimate, the B200 manufacturing-cost figure, and the GB200 NVL72 rack estimate used below; like the GB300 NVL72 figure, treat these as directional market estimates rather than named, dated primary quotes.
DGX B300 System Pricing
NVIDIA positions the DGX B300 as the Blackwell Ultra system built for AI reasoning workloads. The system began shipping in January 2026 and requires mandatory direct liquid cooling, pairing each GPU with 288GB of HBM3e memory – double the standard B200’s capacity. NVIDIA’s own DGX B300 product page is now live on nvidia.com, confirming these specs directly from the manufacturer. As of Q1 2026, and still holding as of August 2026, an 8-GPU DGX B300 system is anchored at $300,000–$350,000, which works out to roughly $37,500–$43,750 per GPU at the system level.
What the System Price Includes
The 8-GPU configuration bundles the GPUs with NVIDIA’s interconnect fabric, system memory, and the integrated software stack. The per-GPU figure above is an effective rate derived from the full $300,000–$350,000 system price – it is not the same as buying GPUs standalone. As of July 2026, a single B300 GPU purchased outright runs about $53,000 (Spheron Network, July 5, 2026), a premium over the single B200‘s $45,000–$50,000, driven by the B300’s larger 288GB HBM3e capacity.
DGX B200 vs DGX B300 System Pricing
With DGX B300 shipments underway, DGX B200 8-GPU systems now have a clearer price anchor too: $280,000–$320,000 as of April 2026 (iFactory App, Thunder Compute), or roughly $35,000–$40,000 per GPU at the system level – compared to the DGX B300’s $300,000–$350,000 ($37,500–$43,750 per GPU) as of Q1 2026, an anchor still holding as of August 2026. The gap reflects the cost of the B300’s jump to 288GB HBM3e memory and the Blackwell Ultra reasoning-optimized platform. The DGX B200 figure is more source-dependent than the B300 figure, though: other 2026 pricing guides place an 8-GPU DGX B200 system as high as $370,000–$500,000, which would erase or even reverse the usual DGX B300 premium depending on configuration. This guide treats iFactory App’s $280,000–$320,000 as the primary DGX B200 anchor since it’s the most specific, consistently-dated quote, but real-world quotes above $320,000 should be expected.
Why DGX B300 Commands a Premium
Blackwell Ultra is designed for AI reasoning – workloads where long chains of inference and large context windows dominate the compute profile. The DGX B300 is NVIDIA’s reference platform for that class of workload, which is why it sits at the top of the Blackwell capex stack rather than competing on raw $/FLOP with the standard B200.
DGX Station: A Desktop Entry Point
Not every DGX B300 buyer needs a rack-mount, 8-GPU system. IntuitionLabs documents a DGX Station – a desktop tower pairing a single Blackwell Ultra B300 GPU with a Grace CPU – priced at $80,000–$125,000 as of August 2026. That’s well below the 8-GPU DGX B300’s $300,000–$350,000 system anchor, giving smaller teams and single-workstation deployments a lower-cost way to access Blackwell Ultra without buying into a full rack-mount system.
B200 Cloud GPU-Hour Pricing
For teams that prefer to rent rather than buy, the B200 is widely available on the cloud. As of August 2026, a broader named-provider comparison spanning 11 platforms puts live B200 GPU-hour rates at $5.91 to $16.11 per GPU-hour – Vast.ai holds the $5.91 floor, Hyperbolic follows at $5.99 and Hyperstack at $6.00, and the range climbs through Lambda ($6.69), Runpod ($6.79), Nebius ($7.15), Vultr ($8.50), and CoreWeave ($8.60) before jumping to hyperscaler pricing at Oracle Cloud ($14.00), AWS ($14.24), and Google Cloud ($16.11), the ceiling. That’s a wider range than the tighter $5.29–$7.05/hr band an earlier August 2026 reading of 6+ specialized providers had shown, since this fuller comparison folds in higher hyperscaler rates the narrower survey didn’t track. It’s still well above the deep spot and promotional pricing – as low as $2.71/hr – that some providers were offering earlier in 2026, reflecting a shift from launch-window promotional rates toward steadier, mainstream cloud pricing as B200 supply has matured.
B200 Cloud Pricing Snapshot
| Metric | Value | Source |
|---|---|---|
| Absolute lowest documented rate | $5.29/hr (Lambda Labs) | GridStackHub August 2026 index |
| Floor (11-provider comparison) | $5.91/hr (Vast.ai) | August 2026 cloud comparison |
| Ceiling (11-provider comparison) | $16.11/hr (Google Cloud) | August 2026 cloud comparison |
| Providers surveyed (fuller comparison) | 11 | August 2026 cloud comparison |
B200 Provider Price Comparison
| Provider | B200 Price (per GPU-hour) | As of |
|---|---|---|
| Vast.ai | $5.91 | August 2026 |
| Hyperbolic | $5.99 | August 2026 |
| Hyperstack | $6.00 | August 2026 |
| Lambda | $6.69 | August 2026 |
| Runpod | $6.79 | August 2026 |
| Nebius | $7.15 | August 2026 |
| Vultr | $8.50 | August 2026 |
| CoreWeave | $8.60 | August 2026 |
| Oracle Cloud | $14.00 | August 2026 |
| AWS | $14.24 | August 2026 |
| Google Cloud | $16.11 | August 2026 |
The Full August 2026 B200 Provider Comparison
The fullest dated look at B200 cloud rates comes from an August 2026 comparison spanning 11 named providers, a broader read than GridStackHub’s earlier August 2026 index of 6+ specialized providers, which had put the market inside a tighter $5.29–$7.05 per GPU-hour band anchored by Lambda Labs’ $5.29 floor – the single lowest B200 rate documented across any named August 2026 source. The wider comparison shows the specialized-provider tier still pricing close to that earlier range – Vast.ai, Hyperbolic, Hyperstack, Lambda, Runpod, and Nebius all fall between $5.91 and $7.15 per GPU-hour – but adds Vultr and CoreWeave at $8.50–$8.60, then a distinct hyperscaler tier at Oracle Cloud, AWS, and Google Cloud running $14.00 to $16.11 per GPU-hour. Earlier in 2026, some providers listed B200 capacity as low as $2.71 per GPU-hour, with hyperscaler-grade listings for smaller GPU configurations running as high as $27.04 per GPU-hour – a roughly 10x spread. The August 2026 comparison narrows that considerably to roughly 2.7x between its $5.91 floor and $16.11 ceiling, suggesting the steepest early promotional discounts have faded even as hyperscaler pricing keeps the top of the range elevated.
Reading the Spread
The gap between the cheapest and most expensive B200 listings depends heavily on which provider tier you compare. As of August 2026, the full 11-provider comparison’s $5.91 floor and $16.11 ceiling put the spread at roughly 2.7x – narrower than the near-10x spread seen earlier in 2026, when promotional floors and hyperscaler ceilings sat much further apart, but wider than the 1.3x spread an earlier, narrower August 2026 reading of 6+ specialized providers had shown before hyperscaler pricing entered the comparison. A single B200 running 24/7 for a month now costs about $4,255 at Vast.ai’s $5.91/hr floor, versus roughly $11,599 at Google Cloud’s $16.11/hr ceiling – a difference of about $7,344 a month. For most buyers, that means provider tier now matters more for raw cost than it did under the narrower specialized-provider-only reading; choosing a specialized neocloud over a hyperscaler is now one of the biggest levers on B200 cost.
B300 Cloud Pricing
B300 availability on the cloud has matured quickly. As of August 2026, Thunder Compute’s pricing table names five on-demand providers running B300 capacity: Modal at $7.10 per GPU-hour, the current floor; Runpod at $7.39; Nebius at $7.85; Oracle Cloud at $15.00; and AWS at $17.80, the ceiling among that named-provider set. A separate August 2026 index adds Spheron’s on-demand B300 SXM6 listing at $9.16 per GPU-hour, slotting in between Nebius and Oracle Cloud, and notes that premium managed DGX B300 stacks – fully managed, turnkey deployments rather than raw on-demand instances – run $12.00 to $18.00 per GPU-hour. That puts the full August 2026 B300 cloud range at $7.10 to $18.00 per GPU-hour once managed offerings are included. Buyers willing to commit to a 48-month reservation can push the effective rate much lower still: one August 2026 pricing guide puts committed, long-term B300 capacity as low as $3.13 per GPU-hour, though that floor applies only to multi-year reserved terms, not on-demand rental. Earlier in 2026, spot and promotional B300 capacity briefly dipped as low as $2.45 per GPU-hour on providers like Spheron Network and Verda – but by August 2026 that short-term deep-discount tier no longer shows up in on-demand named-provider readings, and the effective on-demand floor has moved up to Modal’s $7.10/hr.
B300 Provider Price Comparison
| Provider | B300 Price (per GPU-hour) | As of |
|---|---|---|
| Modal | $7.10 | August 2026 |
| Runpod | $7.39 | August 2026 |
| Nebius | $7.85 | August 2026 |
| Spheron (on-demand SXM6) | $9.16 | August 2026 |
| Oracle Cloud | $15.00 | August 2026 |
| AWS | $17.80 | August 2026 |
B300 Managed and Reserved Pricing
Beyond the named on-demand providers, B300 capacity is also available through two other pricing structures. Premium managed DGX B300 stacks – fully managed, turnkey deployments that bundle support and orchestration on top of the raw GPU – run $12.00 to $18.00 per GPU-hour as of August 2026, the highest tier in this guide. At the opposite end, a 48-month reservation can bring the effective rate down to as low as $3.13 per GPU-hour, though that requires committing to a long-term term rather than paying on demand.
| Tier | Price (per GPU-hour) | Commitment |
|---|---|---|
| Reserved (48-month term) | as low as $3.13 | Multi-year committed capacity |
| On-demand named providers | $7.10–$17.80 | Pay-as-you-go |
| Premium managed DGX B300 stacks | $12.00–$18.00 | Fully managed, turnkey |
What Drives the Provider Spread
Modal’s $7.10/hr floor and Runpod’s $7.39/hr sit close together, reflecting tight competition among serverless and specialized GPU platforms. Nebius prices slightly above that pair at $7.85/hr, and Spheron’s on-demand B300 SXM6 listing lands a bit higher still at $9.16/hr. The bigger jump comes at the hyperscaler and managed tiers: Oracle Cloud’s $15.00/hr and AWS’s $17.80/hr both carry the managed networking, enterprise SLAs, and support overhead typical of major cloud platforms, while fully managed DGX B300 stacks push as high as $18.00/hr for turnkey deployment – roughly double to triple the specialized-provider rates. Teams that don’t need hyperscaler-grade SLAs or fully managed deployment can capture meaningful savings by choosing Modal, Runpod, or Nebius over AWS, Oracle Cloud, or a managed stack, and teams with predictable, long-term demand can go lower still with a 48-month reserved commitment at rates as low as $3.13/hr.
B300 Wholesale and Hyperscaler Pricing (GPUAAS, July 2026)
A July 2026 pricing guide from GPUAAS breaks out B300 pricing from a buyer-type angle rather than a provider angle, and it lines up closely with the named-provider and managed-stack ladder confirmed in August 2026 rather than contradicting it. GPUAAS puts wholesale or committed-volume B300 capacity at $7.00–$12.00 per GPU-hour – the same band where Modal ($7.10), Runpod ($7.39), Nebius ($7.85), and Spheron ($9.16) all price – and pegs hyperscaler on-demand B300 pricing as starting above $12.00 per GPU-hour, the same territory where Oracle Cloud ($15.00), AWS ($17.80), and premium managed DGX B300 stacks (up to $18.00) land. Both figures sit inside the full $7.10–$18.00 per GPU-hour August 2026 range established above, corroborating it rather than widening it.
| Buyer Type | Price (per GPU-hour) | Source |
|---|---|---|
| Wholesale / committed volume | $7.00–$12.00 | GPUAAS, July 2026 |
| Hyperscaler on-demand | $12.00+ | GPUAAS, July 2026 |
B200 vs B300 vs DGX B300 Cost Comparison
Side-by-Side Pricing Summary
| Product | Cloud Low (per GPU-hour) | Cloud High (per GPU-hour) | System-Level Pricing | Reference Date | Source |
|---|---|---|---|---|---|
| B200 (cloud) | $5.91 (Vast.ai) | $16.11 (Google Cloud) | – | August 2026 | 11-provider cloud comparison |
| B300 (cloud) | $7.10 (Modal) | $17.80 (AWS) | – | August 2026 | Thunder Compute |
| DGX B200 (8-GPU system) | – | – | $280,000–$320,000 (~$35,000–$40,000/GPU) | April 2026 | iFactory App, Thunder Compute |
| DGX B300 (8-GPU system) | – | – | $300,000–$350,000 (~$37,500–$43,750/GPU) | Q1 2026 | iFactory App |
| DGX Station (desktop, 1 GPU) | – | – | $80,000–$125,000 | August 2026 | IntuitionLabs |
| GB200 NVL72 (72-GPU rack) | – | – | $2M–$3M (~$27,800–$41,700/GPU) | July 2026 (est.) | market estimates, unconfirmed |
| GB300 NVL72 (72-GPU rack) | – | – | $3M–$4M (~$41,700–$55,600/GPU) | July 2026 (est.) | market estimates, unconfirmed |
Some resellers separately quote fully-configured 8-GPU DGX B300 systems as high as $400,000–$500,000; the table above reflects the more commonly cited OEM/system-level anchor price of $300,000–$350,000, not the high-end reseller quote.
GB200 NVL72: The B200-Based Rack Option
Before Blackwell Ultra, NVIDIA’s rack-scale GB200 NVL72 bundled 72 standard B200 GPUs into a single liquid-cooled rack with full NVLink switching between every GPU. NVIDIA has not published an official price for this configuration either. A July 2026 pricing guide estimates a full GB200 NVL72 rack at roughly $2 million–$3 million, or about $27,800–$41,700 per GPU – below the newer GB300 NVL72’s $3 million–$4 million range, since it carries standard B200 GPUs rather than the higher-memory Blackwell Ultra B300. As with the GB300 NVL72 figure, treat this as a directional estimate rather than a confirmed cost until a named source publishes dated pricing.
GB300 NVL72: The Next Rack-Scale Tier
Above the 8-GPU DGX B300, NVIDIA’s rack-scale GB300 NVL72 bundles 72 Blackwell Ultra GPUs into a single liquid-cooled rack with full NVLink switching between every GPU. NVIDIA has not published an official price. Early market estimates place a full rack at roughly $3 million–$4 million, which works out to about $41,700–$55,600 per GPU – a premium over the DGX B300’s $37,500–$43,750 per GPU system-level rate and over the earlier, B200-based GB200 NVL72’s $2 million–$3 million estimate, reflecting the added NVLink fabric, rack-scale engineering, and the jump to Blackwell Ultra B300 GPUs. Treat this range as a directional estimate rather than a confirmed cost until a named source publishes dated pricing.
Single-GPU Purchase Pricing
Not every buyer wants a full DGX system or a cloud subscription. As of July 2026, a single B200 GPU purchased outright costs about $45,000–$50,000, while a single B300 GPU runs approximately $53,000 (Spheron Network, July 5, 2026) – a premium tied to its 288GB HBM3e capacity. NVIDIA has not officially published retail prices for individual B200 units; the $45,000–$50,000 figure reflects OEM quotes and early listings rather than a manufacturer list price. A separate July 2026 pricing guide quotes a lower B200 “street price” of $30,000–$40,000 per GPU, and pegs NVIDIA’s manufacturing cost for the chip at only about $6,400 – figures that point to a wide margin between production cost and sale price but shouldn’t be read as replacing the OEM-quote range above. These figures exclude networking, cooling infrastructure, and installation. Pricing also varies by reseller: some markets have quoted full DGX B300 systems as high as $400,000–$500,000, well above the $300,000–$350,000 baseline range, depending on configuration and regional markup.
How to Choose
Among the 11 named providers in the August 2026 cloud comparison, B200 is still the cheaper chip to rent on a same-provider basis: at Nebius, Runpod, Oracle Cloud, and AWS – the four providers that list both chips – B300 carries a consistent per-GPU premium of roughly $0.60 to $3.56 an hour over B200. But the overall ranges now overlap. B200’s $5.91–$16.11/hr span (Vast.ai floor to Google Cloud ceiling) reaches well into B300 territory, since hyperscaler B200 pricing from Oracle Cloud, AWS, and Google Cloud runs higher than specialized-provider B300 pricing from Modal, Runpod, and Nebius. In practice, provider choice now matters as much as chip choice: a B200 rented from AWS ($14.24/hr) costs nearly twice as much as a B300 rented from Modal ($7.10/hr). For buyers evaluating capex, the DGX B300 system price of $300,000–$350,000 remains the relevant anchor; for cloud-first teams, choosing the right provider tier – specialized neocloud versus hyperscaler versus managed – now matters at least as much as choosing between B200 and B300. Buyers who only need one or two GPUs rather than a rack-scale system still have a third option: standalone modules at $45,000–$50,000 (B200) or $53,000 (B300) – or, for a preconfigured desktop system, NVIDIA’s DGX Station at $80,000–$125,000 (IntuitionLabs, August 2026). Teams scaling well beyond a single 8-GPU system have two rack-scale tiers to track: the B200-based GB200 NVL72, an early estimate of $2 million–$3 million, and the newer Blackwell Ultra GB300 NVL72, at $3 million–$4 million.
Renting vs. Buying: Where’s the Breakeven?
Every price above answers “how much,” but not “which option is actually cheaper for a given workload.” A simple breakeven comparison – per-GPU system cost divided by cloud GPU-hour rate, assuming continuous 24/7 use – makes the tradeoff concrete. This is a directional estimate, not a full total-cost-of-ownership model: it excludes power, networking, floor space, and staffing on the buy side.
| Comparison | Cloud Rate | Per-GPU System Cost | Approx. Breakeven (24/7 use) |
|---|---|---|---|
| B200 vs Vast.ai floor | $5.91/hr | $35,000–$40,000 | ~8.2–9.4 months |
| B200 vs Google Cloud ceiling | $16.11/hr | $35,000–$40,000 | ~3.0–3.4 months |
| B300 vs Modal floor | $7.10/hr | $37,500–$43,750 | ~7.3–8.6 months |
| B300 vs Nebius | $7.85/hr | $37,500–$43,750 | ~6.6–7.7 months |
| B300 vs Spheron on-demand | $9.16/hr | $37,500–$43,750 | ~5.7–6.6 months |
| B300 vs managed-stack ceiling | $18.00/hr | $37,500–$43,750 | ~2.9–3.4 months |
B200: Rent or Buy?
At Vast.ai’s $5.91/hr floor (August 2026), a rented B200 doesn’t pay back a $280,000–$320,000 DGX B200 system’s per-GPU cost until roughly 8.2–9.4 months of nonstop use – well past the useful window for most short-term projects, which favors renting. At Google Cloud’s $16.11/hr hyperscaler ceiling, that breakeven shortens dramatically to roughly 3.0–3.4 months. That’s a wide gap: unlike the narrower specialized-provider-only reading, the full 11-provider comparison shows the rent-vs-buy math now depends heavily on provider tier as much as deployment length – renting from a specialized neocloud stays cheaper for the better part of a year, while renting the same chip from a hyperscaler can justify buying within a single quarter.
B300: Rent or Buy?
The same pattern holds for B300, shifted slightly by the DGX B300’s higher per-GPU system cost. At Modal’s $7.10/hr floor, breakeven against the $300,000–$350,000 DGX B300 system lands around 7.3–8.6 months. At Nebius’s $7.85/hr, that narrows to roughly 6.6–7.7 months; at Spheron’s $9.16/hr on-demand rate, it narrows further to about 5.7–6.6 months; and at the $18.00/hr premium managed-stack ceiling, breakeven arrives in as little as 2.9–3.4 months. Buyers with predictable, multi-year demand have a fourth path worth pricing out: a 48-month reserved commitment at rates as low as $3.13/hr, which trades upfront commitment length for a dramatically lower effective hourly cost. The practical takeaway: reserve buying outright for workloads with predictable, long-running demand that don’t fit a reservation term, use a 48-month reservation if the workload is both long-running and predictable enough to commit to, and use lower-cost on-demand named providers like Modal or Runpod for shorter or bursty rental needs.
Frequently Asked Questions
How much does a DGX B300 system cost?
As of Q1 2026, an 8-GPU DGX B300 system is anchored at $300,000–$350,000, which implies roughly $37,500–$43,750 per GPU at the system level. That anchor still held as of August 2026, corroborated separately by IntuitionLabs’ figures and by NVIDIA’s own DGX B300 product page, now live on nvidia.com, even though NVIDIA itself still doesn’t publish a system price. The DGX B300 began shipping in January 2026.
How much does the DGX Station cost?
IntuitionLabs documents NVIDIA’s DGX Station – a desktop tower pairing one Blackwell Ultra B300 GPU with a Grace CPU – at $80,000–$125,000 as of August 2026, well below the 8-GPU DGX B300’s $300,000–$350,000 system price. It’s the lowest-cost way to get a preconfigured Blackwell Ultra system directly from NVIDIA’s DGX line.
How much does a DGX B200 system cost compared to DGX B300?
An 8-GPU DGX B200 system is priced at $280,000–$320,000 as of April 2026 (iFactory App, Thunder Compute), while the DGX B300 commands $300,000–$350,000 as of Q1 2026, an anchor still holding as of August 2026. Other 2026 pricing guides place DGX B200 higher still, around $370,000–$500,000, so treat the exact DGX B200 figure as source-dependent rather than fixed. Some resellers have quoted full DGX B300 configurations as high as $400,000–$500,000 depending on market and configuration.
How much does the GB200 NVL72 cost?
NVIDIA has not published an official price for the GB200 NVL72, the B200-based predecessor to the newer Blackwell Ultra GB300 NVL72. A July 2026 pricing guide estimates a full rack of 72 standard B200 GPUs at roughly $2 million–$3 million, or about $27,800–$41,700 per GPU – below the GB300 NVL72’s $3 million–$4 million estimate. Treat this as a directional estimate until a named source confirms dated pricing.
How much does the GB300 NVL72 cost?
NVIDIA has not published an official price for the rack-scale GB300 NVL72, which bundles 72 Blackwell Ultra GPUs into a single liquid-cooled rack. Early market estimates put a full rack at roughly $3 million–$4 million, or about $41,700–$55,600 per GPU – a premium over the DGX B300’s per-GPU system rate and over the earlier B200-based GB200 NVL72 rack, estimated at $2 million–$3 million. Treat this as a directional estimate until a named source confirms dated pricing.
How much does a single B200 or B300 GPU cost to buy?
As of July 2026, a standalone B200 GPU costs about $45,000–$50,000, while a standalone B300 GPU runs approximately $53,000 (Spheron Network, July 5, 2026), reflecting its larger 288GB HBM3e memory. NVIDIA has not officially published either figure as a list price – both are OEM quotes and early-listing estimates. A separate July 2026 pricing guide quotes a lower B200 “street price” of $30,000–$40,000 and estimates NVIDIA’s manufacturing cost at about $6,400 per GPU – figures that hint at wide margins but don’t replace the OEM-quote range above. These prices exclude networking, cooling, and installation.
What is the cheapest B200 cloud rate as of August 2026?
As of August 2026, the single lowest documented B200 GPU-hour rate is $5.29 per GPU-hour at Lambda Labs, per GridStackHub’s August 2026 index of 6+ specialized providers, which puts the narrower market band at $5.29–$7.05 per GPU-hour. Among the fuller 11-named-provider comparison, Vast.ai holds the floor at $5.91 per GPU-hour, with Hyperbolic ($5.99) and Hyperstack ($6.00) close behind; that comparison’s ceiling reaches $16.11 per GPU-hour at Google Cloud, for a $5.91–$16.11 range once hyperscaler pricing from Oracle Cloud, AWS, and Google Cloud is included. That’s still a higher range than earlier in 2026, when some providers offered B200 capacity as low as $2.71 per GPU-hour during launch-window promotions.
How cheap can B300 cloud pricing get?
It depends on the commitment level. Among on-demand named providers, the lowest documented August 2026 B300 rate is Modal at $7.10 per GPU-hour, per Thunder Compute’s pricing table. Runpod follows at $7.39, Nebius at $7.85, Spheron’s on-demand SXM6 listing at $9.16, Oracle Cloud at $15.00, and AWS at $17.80 – with premium managed DGX B300 stacks running as high as $18.00 at the very top. That puts the full August 2026 B300 cloud range at $7.10 to $18.00 per GPU-hour for on-demand and managed capacity. Buyers can go lower by committing to a 48-month reservation, which one August 2026 pricing guide puts as low as $3.13 per GPU-hour – but that rate applies only to multi-year committed capacity, not on-demand rental. GPUAAS’s July 2026 wholesale ($7.00–$12.00/hr) and hyperscaler ($12.00+/hr) bands corroborate this range rather than widening it.
Is B300 always more expensive than B200 in the cloud?
It depends on how you compare. On a same-provider basis, yes: among the four providers that list both chips in the August 2026 data – Nebius, Runpod, Oracle Cloud, and AWS – B300 costs more than B200 every time, by roughly $0.60 to $3.56 per GPU-hour. But the two overall ranges now overlap: B200’s $5.91–$16.11/hr span (Vast.ai to Google Cloud) reaches above B300’s $7.10/hr floor (Modal), since hyperscaler B200 pricing has climbed higher than specialized-provider B300 pricing. A B200 rented from AWS ($14.24/hr) or Google Cloud ($16.11/hr) now costs more than a B300 rented from Modal ($7.10/hr) or Runpod ($7.39/hr). That’s a shift from an earlier, narrower August 2026 reading that had B300 sitting entirely above B200 with no overlap – the picture changed once hyperscaler pricing entered the fuller comparison.
Is it cheaper to rent or buy a Blackwell B200 or B300 GPU?
It depends on how long the workload runs and which provider tier you use. Against the lowest August 2026 cloud rates – Vast.ai’s $5.91/hr for B200, Modal’s $7.10/hr for B300 – renting stays cheaper than buying a DGX system for roughly 7.3–9.4 months of continuous use. Against the highest rates – Google Cloud’s $16.11/hr for B200, or premium managed DGX B300 stacks at up to $18.00/hr – buying pays for itself in as little as 2.9–3.4 months. Workloads predictable and long-running enough to commit to a 48-month reservation can push B300’s effective rate as low as $3.13/hr, changing the math further in favor of renting. As a rule of thumb, short or bursty workloads favor renting from lower-cost specialized providers, hyperscaler or managed rentals make the most sense for short bursts where buying isn’t practical, and predictable, long-running workloads favor either buying a DGX system outright or locking in a multi-year reservation.
When did the DGX B300 start shipping?
The DGX B300 began shipping in January 2026. It requires mandatory direct liquid cooling and pairs each GPU with 288GB of HBM3e memory, double the B200’s capacity.
What workloads is the DGX B300 designed for?
NVIDIA describes the DGX B300 as the Blackwell Ultra system for AI reasoning – workloads characterized by long inference chains and large context windows where the integrated DGX reference architecture provides the most value.
