Scenario wins: laertes (106) hayek-bot (35) Panshul42 (18) cassi (14) Mantic (11) smingers-bot (7)
| Figure/Metric | Value | Source | Significance |
|---|---|---|---|
| Current Lowest Available Price | $4.34/hr | GPUPerHour (June 26, 2026) | The immediate anchor for the forecast. |
| Peak B200 Spot Price | $6.11/hr | Ornn Index (May 30, 2026) | Represents the high-water mark of Blackwell scarcity. |
| NVDA Inventory Growth | +20.5% QoQ | NVIDIA Q1 2027 10-Q | Signals a massive increase in hardware supply. |
| CoreWeave PPE Expansion | $46.6B | CoreWeave Q1 2026 Balance Sheet | Indicates rapid scaling of secondary cloud infrastructure. |
| H100 Post-Peak Spot Rate | $1.49 - $1.99/hr | Market News (May 2026) | Historical floor for previous flagship hardware. |
The GPU spot market historically follows a ‘boom-bust’ cycle linked to NVIDIA’s product cadence. In 2023 and early 2024, the H100 (Hopper) architecture saw spot rates peak near $8.00 per hour due to extreme scarcity. However, as the B200 (Blackwell) began shipping in volume in late 2024 and throughout 2025, H100 spot rates collapsed to a range of $1.49 to $1.99 per hour—a 75% reduction. The B200 is currently following a similar trajectory; it peaked in May 2026 at $6.11 and has already fallen roughly 30% in one month. The transition to the Vera Rubin architecture in Q3 2026 mimics the H100-to-B200 transition, suggesting a sustained downward pressure as Blackwell becomes the ‘previous generation’ hardware. Data from June 2026 shows this pattern is repeating, with price compression accelerating just as next-gen hardware enters production.
My analysis for the NVIDIA B200 SXM spot market is based on a convergence of supply-side expansion and architectural transition. As of late June 2026, the lowest qualifying ‘Available’ secure spot price on gpuperhour.com is $4.34/hr (via Vast.ai). This follows a rapid decline from a May 30 peak of $6.11/hr, indicating that the initial scarcity phase for Blackwell silicon has ended.
Several factors inform the forecast for August 12:
I have weighted a primary scenario (70%) where the price settles between $3.20 and $3.80 as supply continues to outpace demand. A secondary scenario (20%) accounts for price stickiness or temporary stock-outs of the cheapest providers at the exact moment of resolution, which could keep the price near current $4.34 levels. A small tail (10%) accounts for a sharper “compute collapse” similar to the H100’s fall to sub-$2.00 levels, though the B200’s premium status makes a drop below $2.50 unlikely by mid-August.
The resulting prediction reflects a market in a healthy downward trend, nearing the marginal cost of operation for secondary providers as the frontier moves to the Rubin architecture.
Question: what is the lowest GPUPerHour listing for a B200 SXM that is spot/preemptible, secure-tier, in-stock/immediately deployable, and numerically priced at ~2026-08-12 12:00 UTC. Forecasters agree this is a point-in-time market‑microstructure question where the decisive factor is which secure, in‑stock spot/preemptible rows appear on GPUPerHour under the exact filters (Available vs Listed, Deploy action, Secure badge). Current public anchors cluster: a broad GPUPerHour floor near $4.34/GPU‑hr, a strongest plausible secure spot candidate from Nebius at ~$3.95, and clearer secure/public fallbacks like RunPod/Lambda in the ~$5.9–$7/GPU‑hr range; independent aggregators show medians ~ $4.10 and spot minima reported as low as ~$3.95. Sub‑$3 headline examples exist (Spheron, some marketplace/P2P listings, and AWS spot glimpses) but are weakly qualifying for the strict GPUPerHour secure/in‑stock filter and are treated as a lower tail rather than the center. Teams put most probability mass between roughly $3.5–$6/GPU‑hr, with a meaningful left tail if a cheaper qualifying spot row appears and a meaningful right tail if cheap rows are absent/excluded (or if the filter returns zero qualifying rows); supply‑side strength for Blackwell (NVIDIA demand) argues against a large price collapse. Main forecasting guidance: resolve by checking live GPUPerHour filtered rows at snapshot time — the “Available”/Secure status is what will decide whether the ~$3.95 anchors or the $5–$7 fallbacks determine the floor.
Forecast: Probability distribution:
Key drivers are (i) the Secure filter (Tier 3/4), which likely excludes many of the absolute-cheapest community/marketplace offers; (ii) the Spot filter, which excludes providers that only show on-demand pricing (Forecaster 3 notes at least one provider page explicitly saying no spot for B200, implying some visible low prices won’t qualify); and (iii) the “in-stock at the observation minute” requirement, which creates thin-market/availability-churn risk: the cheapest qualifying listing can disappear, causing the minimum to jump to the next provider tier.
Across analysts, the plausible current secure/spot floor spans roughly the mid-$2s to mid-$4s depending on which providers qualify and are in stock; the next tier is commonly around $4–$6 (e.g., RunPod-like secure offerings), with a right tail for moments of broad stock-outs. With ~7 weeks to Aug 12, I expect modest downward drift in the minimum from additional capacity/competition, but not enough to confidently push the secure in-stock floor below ~$2 in most states.
I therefore center the median in the mid-$3s, keep meaningful probability on a ~$2–$3 low end (if at least one low-cost secure spot listing is in stock), and maintain a fat upper tail for temporary scarcity (but still < $14 with very high confidence).
An amateur forecast would commonly (a) anchor on an on-demand B200 price and apply a generic spot discount, (b) ignore that the resolution is the minimum across listings at a single timestamp, and/or (c) miss that the Secure+Spot filters can exclude many seemingly-relevant low prices. My forecast explicitly models those filter/availability mechanics and keeps wider upper-tail uncertainty for stock-out-driven jumps, while still reflecting competitive downward pressure into August. Confidence in improvement: moderate (main residual uncertainty is the exact set of providers classified as “Secure” on gpuperhour.com on the resolution date).
Forecast rationale (numeric):
— Iteration 1 — The forecasts are broadly anchored to recent June 2026 B200 SXM spot listings, especially prices clustered around $2.70–$2.80/hr from major secure providers and aggregators. The main reasoning pattern is that these observed prices likely form the near-term floor if inventory remains available, so the expected August 2026 price is kept close to that level rather than assuming a major re-pricing.
Overall, the collective reasoning suggests that August 12, 2026 lowest advertised spot pricing will most likely stay near the recent ~$2.7–$3.0/hr floor, with outcomes driven primarily by inventory availability, marketplace coverage, and short-term spot volatility.
— Iteration 2 — Across the forecasts, the core reasoning is tightly anchored to recent June 2026 market data, where the lowest observed NVIDIA B200 SXM spot prices were roughly $2.12–$2.74/hour. The dominant expectation is that this range is the best starting point for August, with only modest additional compression by the target date.
Overall, the reasoning converges on a mid-$2/hour floor with wide volatility, driven by recent market observations, filter constraints, and the inherently unstable nature of spot listings.
— Iteration 3 — Across the forecasts, the main reasoning pattern is to anchor on recent mid-June 2026 B200 spot listings and then adjust for how the market may look by August 12 under the specific “In-Stock” and “Secure” filters.
The collective view is that the lowest advertised qualifying B200 SXM spot price on August 12, 2026 is most likely to remain in the low-$2s, perhaps around $2.5–$3.0/hr, with substantial uncertainty driven by market volatility, listing availability, and filter constraints.
Market Dynamics and Downward Pressures Forecasters broadly agree that the B200 spot market is currently experiencing downward price pressure due to increasing supply. This trend is driven by easing manufacturing bottlenecks—specifically improved TSMC yields and memory availability—alongside the impending rollout of next-generation NVIDIA architectures, such as Blackwell Ultra and Vera Rubin. As these newer chips approach deployment in late 2026, secondary cloud providers and sub-lessors are heavily incentivized to aggressively discount and monetize their current B200 capacity to avoid rapid hardware depreciation.
The “Secure” and “In-Stock” Filter Constraints A central point of consensus is the restrictive nature of the specified aggregator filters. The “Secure” requirement mandates enterprise-grade, Tier 3 or Tier 4 data centers. This strictly eliminates ultra-cheap, decentralized, or residential community nodes from the pricing pool, keeping the baseline elevated. Furthermore, the “In-Stock” filter introduces significant upside risk: because automated algorithms and intense demand quickly consume the cheapest nodes, the lowest available spot instance at the exact time of resolution is highly likely to default to the more expensive, established neo-clouds or primary hyperscaler fallback anchors.
Operational Cost Floors and Robust Demand Despite the market’s downward trajectory, rationales emphasize a hard operational price floor dictated by strict data center economics. The high capital expenditure of acquiring B200 hardware, combined with massive power and cooling requirements, means providers cannot viably sustain spot prices below their break-even operational thresholds. Additionally, robust enterprise demand—fueled by major frontier model training and inference cycles, as well as hyperscaler baseline price hikes—creates a “price umbrella” that limits how far independent providers actually need to lower their rates to maintain utilization.
Question: What will be the lowest advertised spot price per-hour for an NVIDIA B200 SXM GPU on August 12, 2026? Final Prediction: Probability distribution:
Total Cost: extra_metadata_in_explanation is disabled Time Spent: extra_metadata_in_explanation is disabled LLMs: extra_metadata_in_explanation is disabled Bot Name: extra_metadata_in_explanation is disabled
Forecaster 1: Probability distribution:
Forecaster 2: Probability distribution:
The research summarizes that NVIDIA B200 SXM spot prices have been volatile in June 2026: a reported peak of $6.11/hour on May 30, 2026 fell to about $4.22/hour by June 21–23, 2026. Marketplace snapshots cited Vast.ai listings in the ~$3.44–$7.45/hr range (median ~$4.5–$4.7/hr), RunPod around ~$5.89/hr, and hyperscaler on‑demand P6-B200 Capacity Block pricing (AWS) listed at $12.355/hr per GPU effective July 1, 2026. Prediction-market activity (Kalshi) was cited as betting that B200 prices would not return to the late‑May peak in the short term; Polymarket data cited low probabilities for related NVDA stock price targets. Reports and provider notices referenced planned contractual/price changes that could affect supply-side pricing: an announced October 2026 contract renewal benchmark of ~$5.10/hr (vs. $2.63 prior) and hyperscaler price increases (AWS +20% July 1). Historical reference classes were used: H100 and H200 generations showed steep post-peak declines (examples: H200 ~40% drop in ~3 weeks; H100 declines of 64–75% from peak over months), spot typically runs 40–70% below on‑demand, and provider/pricing variance can be large. The research also notes ongoing supply constraints, long delivery lead times for bulk orders (12–15 months), emergence of competing Blackwell products (B300), and that gpuperhour.com (the question’s named resolver) was not prominent in the retrieved market-price snapshots, so most pricing evidence came from Vast.ai, RunPod, marketplace aggregators and provider notices.
Sources/websites used (verbatim where available):
I’ll help you gather comprehensive research for this forecasting question about NVIDIA B200 SXM GPU spot pricing. Let me conduct several searches to find relevant news, market trends, base rates, and prediction markets.
Modal: 1.8-3.5 (36.7%) • frontier aggregate • 4m0s
Resolution sources/checks:
Edge cases:
frontier_1 - Modal: 3.5-5.3 (44.0%)
frontier_2 - Modal: 1.8-3.5 (52.0%)
frontier_3 - Modal: 3.5-5.3 (45.0%)
frontier_4 - Modal: 1.8-3.5 (40.0%)
Material notes
Guidance
Revision
| Bin | Probability |
|---|---|
| 0-1.8 | 7.4% |
| 1.8-3.5 | 36.7% |
| 3.5-5.3 | 35.3% |
| 5.3-7 | 14.1% |
| 7-8.8 | 3.7% |
| 8.8-10.5 | 1.3% |
| 10.5-12.3 | 0.8% |
| 12.3-14 | 0.4% |
| Above 14 | 0.3% |
wooley_rhino • run 7a27af58 • $2.62 • 4m0s
Extracted Percentile Values: {0.1: 1.08, 1.0: 1.38, 5.0: 1.82, 10.0: 2.08, 15.0: 2.28, 20.0: 2.45, 25.0: 2.6, 30.0: 2.74, 35.0: 2.86, 40.0: 2.98, 45.0: 3.08, 46.0: 3.1, 47.0: 3.12, 48.0: 3.14, 49.0: 3.16, 50.0: 3.19, 51.0: 3.22, 52.0: 3.25, 53.0: 3.28, 54.0: 3.31, 55.0: 3.35, 60.0: 3.56, 65.0: 3.79, 70.0: 4.05, 75.0: 4.35, 80.0: 4.76, 85.0: 5.3, 90.0: 6.15, 95.0: 7.75, 99.0: 11.6, 99.9: 15.8}
Model’s Answer: Step one is anchoring on the most relevant current evidence rather than broad Blackwell pricing. The best available June 2026 context says raw B200 spot lows across all provider types are roughly 2.12 to 3.70, but the exact resolution filter is narrower: in stock only, spot, secure tier, B200 SXM. The research explicitly notes secure-filtered aggregator listings are often higher, around 3 to 4 plus. That makes the unrestricted 2.12 style lows only partially relevant.
Step two is adjusting for the fact that this resolves on the lowest advertised listing at a single moment, not an average. That pushes the target below the mean secure spot price. If the secure spot average is around the mid-3s today, the instantaneous minimum can plausibly sit in the high-2s to low-3s even without a major market move.
Step three is the time horizon. There are only about seven weeks until resolution. Supply growth and competitive undercutting can move prices meaningfully, but the short window argues against an extreme structural collapse unless secure providers suddenly flood the market. I therefore expect modest softening to be more likely than either a crash below 2 or a squeeze to hyperscaler-like levels.
The main assumptions doing the work here are these. First, that the secure-tier filter will still leave at least a few competitive non-hyperscaler listings at resolution; I think there is about a 30 percent chance this is more wrong than I expect, which would shift the distribution upward. Second, that B200 supply expansion by mid-August is incremental rather than dramatic; I give about a 35 percent chance this assumption is materially wrong, which mostly fattens both tails but especially the low tail. Third, that there will be qualifying listings rather than an ambiguity-triggering zero-listing situation; I think failure risk here is around 10 percent, but I do not let that dominate the numeric quantiles.
A low outcome would come from secure-capable secondary providers aggressively discounting to keep B200s utilized, especially if multi-GPU pod pricing is prorated to an unusually low per-GPU figure. I put roughly a 25 percent chance on resolving below 2.60. A high outcome would come from a thin secure inventory snapshot at exactly resolution time, with only pricier providers in stock or a short-lived demand spike lifting spot asks. I put roughly a 15 percent chance on resolving above 5.30.
Overall, my base case is a resolution near 3.2, with a fairly wide right tail because this is a minimum-over-listings metric on a volatile spot market with thin secure-filter inventory.
Percentile .1: 1.08 Percentile 1: 1.38 Percentile 5: 1.82 Percentile 10: 2.08 Percentile 15: 2.28 Percentile 20: 2.45 Percentile 25: 2.60 Percentile 30: 2.74 Percentile 35: 2.86 Percentile 40: 2.98 Percentile 45: 3.08 Percentile 46: 3.10 Percentile 47: 3.12 Percentile 48: 3.14 Percentile 49: 3.16 Percentile 50: 3.19 Percentile 51: 3.22 Percentile 52: 3.25 Percentile 53: 3.28 Percentile 54: 3.31 Percentile 55: 3.35 Percentile 60: 3.56 Percentile 65: 3.79 Percentile 70: 4.05 Percentile 75: 4.35 Percentile 80: 4.76 Percentile 85: 5.30 Percentile 90: 6.15 Percentile 95: 7.75 Percentile 99: 11.60 Percentile 99.9: 15.80
Forecast (median): $3.9949 per GPU-hour
I treated this as a forecast of the secure-spot floor, not the overall B200 market. GPUPerHour says it updates prices every 60 seconds, spot prices can change multiple times per hour, and the relevant filters include In-Stock, Spot, and Secure versus Community. (gpuperhour.com)
My main current anchor is Nebius, because it is clearly a secure provider and publishes explicit preemptible pricing. Nebius’ current AI Cloud pricing shows NVIDIA HGX B200 at $3.95/GPU-hour preemptible and $7.15 on-demand, and Nebius’ changelog says that during March 16–22, 2026 it added a public single-GPU B200 preset and extended preemptible support to all GPU VM platforms. That matters because the resolution uses the cheapest per-GPU rate, with a single-GPU listing taking precedence when available. (nebius.com)
I then cross-checked other secure providers. GPUPerHour’s provider pages currently show Cirrascale at $4.79/GPU-hour for listed B200 SXM offers, RunPod at available offers of $4.99, $5.89, and $5.89 per GPU, and CoreWeave at $8.60/GPU-hour, while RunPod’s own pricing page shows B200 Pods at $5.89/hour and CoreWeave’s B200 page states that spot instances are available. Taken together, today’s secure market looks like a low-$4s floor with a long right tail when capacity is tight. (gpuperhour.com)
A broader cross-check points in the same direction. Cloud-GPUs’ analytics page, last updated June 23, 2026, shows B200 averages of $6.84 on-demand and $4.10 spot across 13 providers. That is an average, not a floor, but it is close enough to the Nebius $3.95 preemptible anchor to suggest that a secure-spot clearing price around the high-$3s is already realistic today. (cloud-gpus.com)
For forward pressure, I expect mild downward drift rather than a crash. NVIDIA’s official GB300 page says GB300 NVL72 is available now, and Nebius already lists HGX B300 at $4.30 preemptible versus HGX B200 at $3.95 preemptible. That means the next-generation Blackwell Ultra option is only modestly more expensive on at least one major secure cloud, which should cap how far B200 spot prices can stay above the mid-$4 range as more supply appears and providers compete for utilization. (nvidia.com)
My model is therefore a mixture of four scenarios. First, a base case of gradual easing as secure providers compete more aggressively for utilization; second, a flat-to-tight case where demand for frontier training and inference keeps the floor near current levels; third, a stronger discount case where another secure provider undercuts Nebius or RunPod; and fourth, a squeeze case where cheap spot inventory is thin exactly at the August 12, 2026 observation time and the in-stock floor jumps upward. I assign roughly 50%, 25%, 15%, and 10% to those scenarios respectively. This produces a distribution centered a bit below current Nebius pricing, with a median around $3.7/GPU-hour and mean around $3.8/GPU-hour.
My final forecast is that the lowest advertised secure spot price on GPUPerHour at approximately 2026-08-12T12:00:00+00:00 will most likely land in the $3.3 to $4.4 range, with meaningful but smaller probability on a sharper drop toward the high-$2s and a smaller right tail above $5 if in-stock spot capacity is scarce at the snapshot moment. I do note one residual modeling issue from the resolution rules: if zero qualifying listings exist, the market resolves Ambiguous, and that outcome cannot be represented inside the required numeric CDF; the distribution below should therefore be read as conditional on at least one qualifying numeric listing existing.
Resolution will use gpuperhour.com around 2026-08-12 12:00 UTC, filtered to in-stock only, spot, secure, and NVIDIA B200 SXM. GPUPerHour says its prices update every 60 seconds, and the comparison page exposes the exact relevant filters: Show In-Stock Only, Pricing Type, Security Tier, and GPU Count. (gpuperhour.com)
My main current anchor is official Nebius pricing, because it is both secure/enterprise and explicitly publishes preemptible B200 rates. Nebius lists NVIDIA B200 preemptible at $3.95 per GPU-hour from June 1, 2026, versus $2.90 before June 1; standard B200 is $7.15 per GPU-hour. Nebius also already has self-service B300 live, with preemptible B300 at $4.30 per GPU-hour. Outside Nebius, official secure B200 pricing is still materially higher: RunPod lists B200 in Secure Cloud at $5.89 per hour, and Lambda lists self-serve B200 instances at $6.69 per GPU-hour. A broader non-P2P market snapshot from Cloud GPUs shows median B200 spot pricing of $4.10 across 13 providers as of June 26, 2026. (docs.nebius.com)
The main reason I am not simply forecasting something near $3.95 is that newer supply is clearly arriving. Nebius already offers self-service B300, Hyperstack is advertising B300 arrival in Q3/Q4 2026, and NVIDIA said Rubin-based products will be available from partners in the second half of 2026, with cloud adopters including CoreWeave, Lambda, and Nebius. My inference is that even before Rubin is widely rentable on the exact resolution page, the competitive pressure from Blackwell Ultra and Rubin-era supply should lean B200 prices downward into August 12, 2026. (nebius.com)
The counterargument is that B200 pricing has not been falling in a smooth straight line. Nebius raised its own official B200 preemptible rate from $2.90 to $3.95 on June 1, 2026, and official secure B200 prices at RunPod and Lambda remain well above Nebius. That makes a very fast collapse into the low $2s possible, but not the base case. It also means a tight snapshot with only a few secure spot listings could still resolve above $4.50. Because the market is measured at a single instant rather than averaged over a day, I want a fairly fat right tail. (docs.nebius.com)
I therefore model this as a three-scenario mixture: roughly 50% status-quo competition with Nebius-like pricing still setting the floor, 35% chance that another secure provider exposes more aggressive B200 spot pricing by mid-August and pushes the floor lower, and 15% chance of a tighter or noisier market snapshot that leaves the floor meaningfully higher. That produces a median forecast of $3.80 per GPU-hour and a mean just under $3.9. In the CDF, I encode about 20% probability below $3.20, 50% below $3.80, 80% below $4.60, and 95% below $6.00. I judge ambiguous non-numeric resolution risk to be low enough not to encode separately in the numeric CDF, because there are already multiple current B200 listings and the rules specify fallback sources if GPUPerHour is unavailable. (docs.nebius.com)
The resolution source is gpuperhour.com with In-Stock Only ON, Pricing Type Spot, Security Tier Secure, and GPU Model NVIDIA B200 SXM. GPUPerHour explicitly separates Secure from Community listings, and its own explanation says peer-to-peer/community marketplaces such as Vast.ai tend to be the cheapest but have fewer enterprise guarantees. That matters here because many headline-low B200 prices come from community inventory and should not be used as the main anchor for a secure-only forecast. (gpuperhour.com)
For the current secure spot floor, the clearest official anchor is Nebius. Nebius’s official pricing page lists NVIDIA HGX B200 at $3.95 per GPU-hour for preemptible capacity and $7.15 on-demand. CoreWeave’s official pricing page lists NVIDIA HGX B200 at $34.87 per hour spot on an 8-GPU node, which is $4.35875 per GPU-hour. Those two prices bracket the most credible current secure-spot range at roughly the high-$3s to low-$4s. (nebius.com)
Other secure providers mostly give higher or non-spot B200 signals today. Runpod’s pricing page shows B200 Secure Cloud at $5.89/hr; Lambda’s self-serve B200 instance pricing is $6.69-$6.89/hr and Lambda’s GPUPerHour page says B200 has no spot pricing; VERDA’s GPUPerHour page says B200 is on-demand only; and Cirrascale’s official pricing is term/month-based with hourly equivalents from $4.79 to $5.99 per GPU-hour. I treat these as competitive pressure and upper-range context, but not as the most likely resolving low spot print. (runpod.io)
A broader market cross-check points the same way. Cloud-gpus.com’s June 26, 2026 snapshot shows a B200 spot median of $4.10 across 13 providers. GPUPerHour’s B200 rental page recently showed 69 offers from 9 providers with 10 in-stock instances across 16 regions, but that broader page also includes community and other non-resolving inventory, so I use it mostly as a breadth/availability check rather than the exact resolving anchor. (cloud-gpus.com)
Looking ahead to August 12, 2026, the main bearish force is upcoming supply expansion and substitution risk. NVIDIA said DGX Cloud Lepton connects developers with tens of thousands of GPUs from partners including CoreWeave, Lambda, Nebius and others, and on May 31, 2026 NVIDIA said Vera Rubin was ramping into full production across 350+ factories in 30 countries. Hyperstack is already advertising B300 early access in Q3/Q4 2026. My inference is that even before Rubin/B300 dominate deployed capacity, the neocloud market will price in the next wave and push sellers to discount B200 spot inventory more aggressively. (nvidianews.nvidia.com)
The main bullish counterweight is time: August 12, 2026 is only 47 days after June 26, 2026. NVIDIA’s May 20, 2026 earnings release described AI-factory buildout as accelerating and reported record data-center revenue, which argues that demand for frontier GPUs is still extremely strong. So I do not expect a dramatic collapse in secure B200 spot pricing over such a short horizon; I expect modest downward drift rather than a step-change lower. (investor.nvidia.com)
My quantitative forecast is a three-scenario mixture. Scenario 1 (25%) is aggressive competitive repricing or a new secure-spot entrant, centered near $3.15. Scenario 2 (60%) is mild downward drift from today’s Nebius-like floor, centered near $3.75. Scenario 3 (15%) is tighter availability, centered near $4.55, where Nebius is unavailable and a CoreWeave-like price becomes the floor. The weighted mean of that mixture is $3.72/GPU-hour. (nebius.com)
I assign only a small chance to an Ambiguous resolution because there are already multiple enterprise providers with live B200 inventory and pricing today; for the required numeric CDF I am effectively conditioning on a numeric resolution. Overall, my median is in the mid-$3.7s and my point estimate for the lowest advertised secure spot price on August 12, 2026 is $3.72 per GPU-hour. (gpuperhour.com)
My best single-number estimate is $4.25 per GPU-hour. The late-June 2026 market is already showing a wide but not collapsed B200 SXM range: GPUPerHour pages and snippets show Hyperstack around $3.50, Nebius at $3.95, Cirrascale at $4.79, RunPod around $5.89, VERDA around $4.89-$6.11, Lambda Labs at $6.99, and CoreWeave at $8.60, while the generic B200 page crawled on June 24, 2026 highlighted VERDA at $6.11 as the cheapest in-stock option and 68-69 live offers overall. (gpuperhour.com)
The biggest forecasting issue is that GPUPerHour’s own pages are internally inconsistent in ways that matter for resolution. Nebius shows 1 offer available at $3.95 but with status Listed, not Available; VERDA’s provider page shows $4.89 but all 20 offers are sold out; yet a June 24 crawl of the generic rent page surfaced a lowest in-stock price of $6.11 on VERDA. Because the question resolves from a single noon-UTC snapshot with in-stock, secure, and spot filters, the decisive variable is whether one low-price secure listing is actually live at that instant, not just the lowest published rate somewhere in the ecosystem. That pushes me toward a wider-than-usual distribution. (gpuperhour.com)
For direction over the next 47 days, I lean modestly downward rather than upward. Nebius officially raised B200 preemptible pricing on June 1, 2026 from $2.90 to $3.95, so this is not a market already in free fall. But B300 is already rentable on RunPod at $7.39 and on VERDA at $6.99, GPUPerHour currently tracks far more B200 offers than B300 offers (304 vs. 49), Hyperstack is advertising B300 arrival in Q3/Q4 2026, and NVIDIA said Vera Rubin is ramping into full production with first cloud deployments in 2026. My inference is that this creates real substitution pressure on B200 by mid-August 2026, but probably not enough time for a full repricing of the secure spot market. (docs.nebius.com)
A rough cost-floor check also argues against an ultra-bearish August print. Public estimates still put B200 sale prices around $30,000-$40,000 per chip, and even a simple amortization of $60,000-$80,000 installed cost per GPU over 3 years at 70% utilization yields about $3.26-$4.35 per utilized GPU-hour before power and operating overhead. Spot pricing can of course dip below full cost for short windows, but sustained secure-datacenter B200 pricing much below the low-$2s would imply very deep negative gross margins for enterprise providers. That can happen in isolated distressed listings, but it does not look like the base case for a secure filter. (epoch.ai)
So I use a three-scenario mixture with a tiny residual extreme right tail. About 28% weight goes to a bearish scenario where a secure low-end provider such as Nebius, Hyperstack, or a newly aggressive competitor has live capacity and the resolved price lands around $2.5-$3.5. About 52% goes to a middle scenario where competitive pressure lowers the current effective floor but the exact filtered noon snapshot still lands around $3.8-$5.2. About 19.5% goes to a tighter-market scenario where secure in-stock capacity remains scarce and the printed low is around $5.5-$7.0. I leave a 0.5% far-right residual tail for unusual market or listing conditions. That produces a median around $4.25 and a mean a bit above $4.3, which is my final forecast. (gpuperhour.com)
August 12, 2026 is only 47 days away, so this is mainly a near-term market-structure forecast. The key resolution detail is that the question resolves to the lowest immediately deployable secure spot listing at one snapshot, not an average market price, so temporary stockouts matter a lot.
Today’s best hard price anchor is Nebius. Its official AI Cloud pricing page lists NVIDIA B200 NVLink at $5.50 per GPU-hour on-demand and $2.90 preemptible, while GPUPerHour’s Nebius B200 page says B200 spot/preemptible exists and currently shows a listed B200 offer. Because one live GPUPerHour Nebius page shows $3.95 while Nebius’s own pricing page shows $2.90 preemptible and $5.50 on-demand, I treat the current secure-spot floor as being in the low-$3s rather than pretending the exact present print is perfectly observed. (docs.nebius.com)
The rest of the secure field provides the right side of the near-term distribution. Runpod’s official secure-cloud pricing page lists B200 at $5.89/hr, says spot/preemptible GPUs are available on the platform, and GPUPerHour’s RunPod provider page lists B200 at $4.99 as RunPod’s current lowest B200 rate. CoreWeave’s GPUPerHour provider page lists B200 SXM at $8.60 and says spot instances are available. AWS also offers spot instances generally on GPUPerHour, though its current GPUPerHour inventory page does not presently show a B200 listing. If the cheapest Nebius-like listing is absent exactly at the resolution snapshot, the floor can jump quickly to the next provider tier. (runpod.io)
Just as important, several enterprise B200 providers do not currently offer spot for this GPU: Cirrascale, Hyperstack, Lambda Labs, VERDA, and Vultr all say B200 is on-demand-only on their current GPUPerHour pages. That narrows the likely competition set for the resolving price and makes the minimum listing more sensitive to a small number of vendors than a casual read of the whole B200 market would suggest. (gpuperhour.com)
There is real downward pressure on B200 pricing, but it is gradual rather than overwhelming for this exact date. CoreWeave said HGX B300 became generally available on March 16, 2026. Runpod’s official pricing page now lists B300 at $7.39/hr, and Nebius’s official pricing page lists B300 preemptible at $3.40/hr. NVIDIA says Rubin-based products will be available from partners in the second half of 2026, with CoreWeave, Lambda, and Nebius among the first cloud providers. Since August 12, 2026 falls early in H2, I treat Rubin as a mild price-depressing force by the resolution date, but not as something certain to have fully commoditized B200 yet. (coreweave.com)
The opposing force is that frontier supply still looks tight. Runpod wrote that teams needing H100, H200, B200, or B300 class hardware still need to plan weeks or months ahead, even while Blackwell supply is projected to nearly quadruple by mid-2026 and Runpod is adding thousands of GPUs weekly. Nebius’s February 2026 shareholder letter says it is scaling toward 800 MW to 1 GW of connected power by year-end, but it also highlights multi-year deals with Microsoft and Meta. My read is that supply is expanding, yet a meaningful fraction of the best capacity can still be absorbed by large commitments instead of flowing into the cheapest spot inventory. (runpod.io)
Volatility also matters. GPUPerHour says it updates prices every 60 seconds and that spot prices can change multiple times per hour. Its current Cloud GPU Models page shows B200 still priced well above H200 across tracked offers at $6.11 versus $3.50, and below B300 at $7.50, which argues against both an immediate collapse to H200-like pricing and a reversion to hyperscaler-like highs for the very cheapest secure spot listing. Separately, ORNN data summarized by Tomasz Tunguz showed B200 spot moving from $2.31 in early March to $4.95 in late April 2026, a reminder that Blackwell spot pricing can move sharply on demand shocks. (gpuperhour.com)
Putting those pieces together, my base case is that the August 12 floor is still set by a Nebius-like or similarly aggressive secure spot listing, with mild downward pressure from B300 substitution but not enough time for a full collapse. I therefore center the forecast at a median of about $3.06 per GPU-hour. Most of my probability mass sits between roughly $2.4 and $4.8, with a meaningful right tail above $5 if the cheapest vendor is unavailable at the snapshot or if Blackwell demand tightens again. I assign only a very small chance to an ambiguous or non-numeric outcome because multiple secure providers with spot support are live today and the question includes fallback aggregation sources. (gpuperhour.com)