NVIDIA’s B200 accelerator, the chip that anchors most large-scale AI training and inference clusters built in the past year, is now worth more on the resale market than it cost when it first shipped. New data published on September 16, 2026 shows the B200’s residual value has climbed to 158 percent of its original launch price, a 58 percent premium that has stunned even veteran hardware analysts. The finding, first reported by Wccftech citing figures from market-data firm Silicon Data, flips the basic assumption that underpins nearly every corporate IT budget: that compute hardware depreciates, not appreciates.
For a piece of silicon that has been in mass availability for roughly a year, a 58 percent premium over launch price is not a rounding error. It is a market signal. It tells buyers, lessors, and cloud providers that the AI compute shortage most executives assumed would ease by now has instead tightened, and it raises a question every CFO evaluating an AI infrastructure budget in September 2026 needs answered: are you buying an asset, or renting access to one that is actively becoming scarcer?
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What the 158 Percent Residual Value Number Actually Means
Residual value is an accounting concept, not a marketing one. Under a standard straight-line depreciation schedule, a piece of equipment loses value evenly over its estimated useful life. A $1,000 asset with a five-year useful life loses $200 in book value each year; after two years, its residual value would sit at $600. That is how enterprise finance teams, leasing companies, and cloud providers typically model GPU fleets when they build multi-year total cost of ownership projections.
NVIDIA’s B200 is defying that model entirely. According to Silicon Data’s analysis, cited by Wccftech’s Rohail Saleem on September 16, 2026, the B200’s residual value today is 158 percent of its original cost, meaning the chip is trading at a 58 percent premium to what buyers paid for it roughly a year ago. Silicon Data’s research also found that NVIDIA’s older A100 and H100 GPUs are commanding residual values “well in excess of what 3 or 5-year straight-line schedules would imply,” extending the pattern beyond the newest Blackwell generation into hardware that is now several architecture cycles old.
Silicon Data flagged the finding in a post on X (formerly Twitter) on September 16, 2026, framing it alongside a broader observation about AI infrastructure financing: startups reserving compute are increasingly being asked to commit to three-year terms with 30 to 40 percent upfront payments, rather than the one-year terms that were common earlier in the AI buildout cycle. That detail matters because it shows compute providers are not padding contracts arbitrarily. They are responding to a hardware supply chain that cannot keep pace with demand.
The Rental Market Tells the Same Story
The residual value spike is not an isolated data point. Wccftech’s report also referenced GPU rental pricing that Silicon Data compiled earlier this year: in January 2026, B200 hourly rental rates sat just below the $5 per hour threshold. By August 2026, those rates had climbed to between $5.50 and $5.80 per hour, an appreciation of 50 to 80 percent in roughly seven months. When both the purchase-resale market and the hourly rental market move in the same direction at the same time, it is a strong indication that the underlying constraint is physical supply, not speculative trading.
Cloud GPU pricing trackers add further texture to that picture, though figures vary widely by provider and contract structure. Neo-cloud providers were quoting a mean of roughly $5.09 per hour for B200 instances in Q2 2026, while on-demand rates at major hyperscalers like AWS could reach as high as $14.24 per GPU-hour. Spot pricing at specialist providers, by contrast, has dipped as low as $2.12 per hour. That spread, from roughly $2 to over $14 an hour for the same underlying chip, reflects a market where availability, not just raw compute performance, is what buyers are really paying for.
Why the B200 Refuses to Depreciate
Three forces are converging to keep the B200 valuable well past the point where older hardware generations would normally have started sliding down the depreciation curve.
The first is inference economics. The B200 has proven unusually efficient at serving large language model inference workloads, not just training them. Wccftech’s report cites SemiAnalysis estimates that put B200 inference costs at just $0.20 per 1 million tokens, running at 76 tokens per second per user, on DeepSeek R1. Inference now consumes a growing share of total AI compute spend as more companies move from experimenting with models to running them in production at scale, and a chip that keeps token costs low has a durable reason to stay in demand long after training clusters have moved on to newer silicon.
The second is scarcity. NVIDIA’s Blackwell architecture, which underpins the B200 and DGX B200 platform, has been supply-constrained since launch. Advanced packaging capacity, HBM3e and HBM4 memory availability, and CoWoS interconnect capacity at TSMC have all acted as bottlenecks on how many Blackwell-class chips NVIDIA can actually ship in a given quarter, regardless of how strong demand is. Those two inputs are not minor line items, either: Epoch AI’s analysis, as reported by Axis Intelligence in December 2025, found that HBM3e and packaging alone account for roughly two-thirds of the B200’s total manufacturing cost, which helps explain why packaging and memory capacity, not logic-die production, have become the binding constraint on supply.
The third is the AI buildout itself, which shows no sign of slowing despite public calls for a more measured pace from some of the industry’s most prominent voices. Wccftech’s report specifically notes that the current upcycle has continued “despite calls for pacing the development cadence of AI models from the likes of Anthropic’s Dario Amodei and OpenAI’s Sam Altman.” In other words, even executives at the labs building the most capable models have publicly floated the idea of slowing down, and the market has not listened. Compute demand keeps climbing anyway.
B200 Pricing: List, Street, and Cloud Rates Compared
Multiple industry pricing trackers, compiled through September 2026, illustrate just how wide the gap has become between what NVIDIA nominally charges for Blackwell hardware and what buyers are actually paying once supply constraints and reseller premiums are factored in. The trend line has barely budged in over a year: Modal Labs pegged B200 SXM pricing at $30,000 to $40,000 per GPU back in July 2025, IntuitionLabs put sale prices at $30,000 to $50,000 per GPU that December, AI Wiki still had unit pricing around $30,000 to $40,000 as of May 2026, and GPU Smith’s July 2026 street-pricing estimate landed in that same $30,000-to-$40,000 band, before IntuitionLabs’ most recent September 2026 analysis pegged fully integrated 8-GPU HGX B200 systems at $400,000 to $500,000.
| Pricing Tier | Approximate Range | Notes |
|---|---|---|
| B200 SXM, 8-GPU volume list price | $30,000 – $40,000 per GPU | NVIDIA’s nominal per-unit price in bulk HGX configurations |
| B200 SXM, street/spot quote | $45,000 – $55,000 per GPU | Actual resale/reseller pricing under 2026 supply constraints |
| 8-GPU HGX B200 server | $400,000 – $500,000 | Typical integrated quote around $450,000 |
| DGX B200 appliance | ~$515,000 | Fully integrated 8-GPU NVIDIA appliance |
| Cloud rental, spot rate | $2.12 – $2.74 / hour | Specialist and neo-cloud providers |
| Cloud rental, on-demand (neo-cloud) | $3.70 – $6.04 / hour | Mean around $5.09/hr in Q2 2026 per provider trackers |
| Cloud rental, on-demand (hyperscaler) | Up to $14.24 / hour | AWS Blackwell instance list rate |
That spread between the roughly $30,000 list price NVIDIA quotes in volume and the $45,000-plus street price buyers are actually paying is, in effect, the residual-value story playing out in real time on the primary market, not just the secondary one. It looks even starker next to Epoch AI’s December 2025 estimate that actual manufacturing cost for a B200 runs just $5,700 to $7,300 per unit, meaning street buyers are paying somewhere between six and nine times the chip’s underlying production cost just to get it delivered on a reasonable timeline.
How B200 Stacks Up Against H100 and H200
The B200’s premium becomes even more striking when set against NVIDIA’s prior-generation Hopper chips, which were themselves considered scarce and expensive not long ago. Pricing trackers through mid-2026 show Hopper-class hardware has not gotten meaningfully cheaper even as Blackwell has ramped, which undercuts the usual assumption that a new architecture generation pulls older stock’s value down.
| GPU | Architecture | Typical Cloud On-Demand Rate | Market Behavior in 2026 |
|---|---|---|---|
| H100 | Hopper | ~$2.01 – $2.53 / hour | Residual value exceeding straight-line depreciation estimates |
| H200 | Hopper (refresh) | Contract pricing up ~40% (Oct 2025 – Mar 2026) | Roughly half of tracked providers reported no capacity coming off contract |
| B200 | Blackwell | ~$3.70 – $6.04 / hour (neo-cloud) | 158% residual value; 58% premium over launch price |
The H200 data point is particularly telling. Contract pricing for H200 capacity rose roughly 40 percent between October 2025 and March 2026, even as B200 supply was supposedly ramping to take pressure off the market. Instead, roughly half of tracked specialist cloud providers reported no Hopper-class capacity coming off contract at all during that window, which suggests demand for AI compute broadly, not just for the newest chip, has continued to outstrip supply across NVIDIA’s last two full architecture generations simultaneously.
Historical Context: GPUs Weren’t Supposed to Work This Way
For most of the GPU industry’s history, the depreciation curve was predictable and steep. Consumer and even professional-grade graphics cards typically lost 20 to 40 percent of their value within the first year, and data center accelerators followed a similar pattern once a successor architecture shipped. NVIDIA’s own V100, launched in 2017, and the A100 that followed it in 2020, both depreciated in line with conventional IT hardware expectations for years, even amid periodic shortages tied to cryptocurrency mining demand.
What changed was the emergence of large language models as a mainstream commercial product starting in late 2022, followed by a compute buildout that has now run for roughly four consecutive years without a meaningful correction. Silicon Data’s finding that even the A100, a chip that first shipped more than five years ago, is holding value “well in excess” of straight-line depreciation schedules is arguably the more remarkable data point buried inside the B200 headline. It suggests the entire NVIDIA data center GPU stock, not just the newest flagship, has been repriced by the market as a scarce, appreciating asset class rather than a depreciating capital expenditure.
Market Impact: What This Means for AI Infrastructure Spending
The practical consequences of a GPU that appreciates rather than depreciates ripple through nearly every layer of the AI economy.
For cloud providers and neoclouds, a GPU fleet that holds or gains value changes the collateral math on debt-financed data center buildouts. Lenders have historically discounted GPU-backed loans aggressively because hardware depreciates fast; a chip that is worth more a year later is, in theory, better collateral, which could make it easier for infrastructure operators to raise debt against existing fleets rather than pure equity.
For AI startups, the financing terms described by Silicon Data, three-year compute reservations with 30 to 40 percent paid upfront, represent a meaningful working-capital burden that did not exist to the same degree a year or two ago. Founders who once could reserve a year of compute on relatively light terms are now being asked to lock in multi-year commitments before they know whether their product-market fit will hold that long.
For enterprises building internal AI infrastructure, the case for buying rather than renting gets more complicated, not less. A depreciating asset is easy to model on a balance sheet. An appreciating one raises awkward questions: should finance teams treat GPU purchases more like real estate or fine art than like laptops, and does that change how boards evaluate capital expenditure requests for AI infrastructure going forward?
For NVIDIA itself, sustained scarcity across three consecutive architecture generations, Hopper, Blackwell, and the upcoming Vera Rubin platform, reinforces the company’s pricing power heading into its next product cycles, even as competitors like AMD’s MI350-series accelerators gain incremental market share in specific workloads.
Competitive Landscape: Where AMD and the Hyperscalers Fit In
NVIDIA’s residual-value story does not exist in a vacuum. AMD has continued pushing its Instinct MI-series accelerators as a lower-cost alternative for specific training and inference workloads, and hyperscalers including Google, Amazon, and Microsoft have all continued expanding their own custom silicon programs, TPUs, Trainium, and Maia chips respectively, specifically to reduce their dependence on NVIDIA allocations. None of those alternatives has yet dented NVIDIA’s core data center market share in a way that shows up in the residual value data, but the fact that hyperscalers are willing to spend billions developing competing silicon, rather than simply paying NVIDIA’s prices, underscores just how costly the current scarcity has become for the buyers footing the bill.
The competitive pressure is also visible in secondary markets outside the US. Reports of modified, higher-memory versions of NVIDIA consumer and prosumer GPUs circulating through resellers in markets like China point to the same underlying dynamic: demand for NVIDIA silicon, in any form that can be acquired, consistently outpaces what official channels can supply.
What Silicon Data’s Methodology Suggests
Silicon Data’s residual value calculation follows the standard accounting approach: original cost minus cumulative depreciation equals residual value. What makes the B200 (and A100/H100) figures unusual is not the formula but the input, because the market-observed resale and rental prices used to back into that residual value are running well above what any conventional depreciation schedule would predict. That is effectively a bottom-up, market-price-driven audit of how much these chips are actually worth today, as opposed to what a finance department’s amortization table says they should be worth.
It is worth noting that this kind of analysis is inherently a snapshot. GPU markets have moved quickly before, and a wave of new Blackwell Ultra or Vera Rubin supply hitting the market could, in principle, soften B200 resale values within a matter of quarters. But for now, a year after mass availability, the data points in one direction: buyers still cannot get enough of this chip.
Predictions: Where the B200 Market Goes From Here
- Premium pricing persists through at least early 2027. With CoWoS packaging and HBM memory capacity still the binding constraints on Blackwell supply, most analysts tracking the GPU pricing indexes referenced above do not expect a meaningful supply correction before Vera Rubin-class hardware reaches volume production.
- Financing terms keep tightening. If three-year, 30-to-40-percent-upfront compute reservations are becoming the norm rather than the exception, expect more AI startups to explore GPU-backed debt, compute co-ops, or fractional ownership structures to spread that burden.
- Hopper-class hardware stays valuable longer than expected. The fact that A100 and H100 residual values already exceed straight-line projections suggests these chips will remain in active production service well past the point most IT depreciation schedules assumed they would be retired.
- Secondary and gray-market channels expand. As official allocation remains constrained, expect continued growth in reseller and modified-hardware channels, particularly in markets facing export restrictions on the newest NVIDIA silicon.
- Inference efficiency becomes the next pricing battleground. With B200’s low per-token inference cost cited as a key driver of demand, expect NVIDIA’s future architecture announcements to lean heavily on inference cost-per-token benchmarks rather than raw training FLOPS, since that is increasingly what is driving real-world purchasing decisions.
What Buyers Should Actually Do With This Information
For organizations currently weighing whether to buy, lease, or rent B200 capacity, the residual value data cuts in a specific direction: hardware purchased today is unlikely to be a sunk cost a year from now, which changes the calculus around committing capital versus committing to long-term cloud contracts. That said, the same scarcity that is propping up resale values also means lead times for new orders remain long, and buyers negotiating fresh allocations should expect the multi-year, upfront-heavy financing terms described by Silicon Data to be a starting point for negotiation, not an outlier.
Enterprises evaluating rental versus ownership should also factor in the rental rate trend directly: a jump from under $5 per hour in January to $5.50-$5.80 by August is a meaningfully different cost structure over a multi-year AI project than what most 2025-era budget models assumed. Locking in longer-term rental agreements now, before rates climb further, may be more attractive than waiting for supply to loosen.
Frequently Asked Questions
What is the NVIDIA B200’s current residual value?
According to Silicon Data figures reported by Wccftech on September 16, 2026, the B200’s residual value is 158 percent of its original launch price, a 58 percent premium, roughly one year after mass availability began.
Why is the B200 worth more than its launch price?
Reports point to a combination of continued scarcity in Blackwell architecture supply, strong demand for low-cost AI inference (with SemiAnalysis estimating B200 inference costs around $0.20 per 1 million tokens on DeepSeek R1), and an AI compute buildout that has not slowed despite public commentary from industry leaders about pacing AI development.
How much does an NVIDIA B200 cost to buy today?
Industry pricing trackers put NVIDIA’s list price for B200 units at roughly $30,000 to $40,000 per GPU in 8-GPU volume configurations, a range that has held remarkably steady since Modal Labs first clocked it in July 2025, while actual street quotes under 2026 supply constraints have run $45,000 to $55,000 per unit, far above Epoch AI’s December 2025 estimate that the chip costs just $5,700 to $7,300 to actually manufacture.
What does a full DGX B200 or HGX B200 server cost?
A fully integrated DGX B200 appliance lists at approximately $515,000. An 8-GPU HGX B200 server typically runs $400,000 to $500,000, with integrated quotes commonly landing around $450,000.
How much does it cost to rent a B200 GPU in the cloud?
Rates vary widely by provider. Spot pricing at specialist providers has run as low as $2.12 per hour, neo-cloud on-demand rates have averaged around $5.09 per hour in Q2 2026, and hyperscaler list rates such as AWS’s Blackwell instance pricing can reach $14.24 per GPU-hour.
Are older NVIDIA GPUs like the H100 and A100 also holding value?
Yes. Silicon Data’s research found that both the A100 and H100 currently carry residual values well above what standard 3-to-5-year straight-line depreciation schedules would predict, indicating the value retention pattern extends beyond the newest Blackwell generation.
Who is Silicon Data and where did this data come from?
Silicon Data is a market-data research firm that tracks AI infrastructure pricing and published its residual value findings in a post on X on September 16, 2026. Wccftech’s Rohail Saleem first reported the figures in detail the same day.
Does this mean the AI GPU shortage will get worse before it gets better?
Current pricing and residual value trends suggest supply constraints are likely to persist at least into early 2027, based on the packaging and memory bottlenecks cited across multiple 2026 industry pricing reports, though a new architecture generation reaching volume production could eventually ease pressure on Blackwell-class pricing.
Related Coverage
- Nvidia B200's 58% Premium Fuels GPU Loan Boom [2026]
- RTX 5090 Hits $4,329 as Nvidia Delays RTX 60 to 2028 [2026]
- AMD Overtakes Nvidia in GPU Sales as RTX 5090 Hits $9,500 [2026]
- Apple, Nvidia NVLink Fusion Talks for M8 Ultra Server [2026]
- Nvidia Control Panel Settings: 12 Steps to More FPS [2026]
