GPU Tech Refresh: When to Upgrade Your AI Cluster

To decide when to refresh and upgrade your GPU cluster's hardware, you need to consider three things: future demand, the opportunity cost, and the capital required. These three things are tricky because, unexpectedly, instead of falling further, H100 1-year rental rates have rebounded 40%, from $1.70/hr in October to $2.35/hr in March 2026. ^{[1]SemiAnalysis, "The Great GPU Shortage: Rental Capacity, H100 1 Year Rental Price Index" (April 2026)} Understanding future demand and opportunity costs is now a lot harder for cluster owners.

Three options

• Full refresh: Sell the entire fleet, and buy the next generation.
• Partial refresh: Keep a portion of the fleet and add newer hardware alongside it.
• Hold: Keep operating the H100 fleet.

The comparison below models all three for a 256-GPU H100 SXM5 fleet purchased in mid-2023, with the loan approaching maturity, and a potential upgrade to B200s. We use B200s as the modeling proxy because pricing and utilization data is available. In practice, the upgrade target could be B300s, GB-NVL72 racks, or whatever is shipping when you pull the trigger.

At the default 80% utilization, full refresh produces the highest 3-year net profit because B200 rates are roughly double H100 rates. Below ~70% utilization, hold becomes the better option: the H100 fleet requires no new capital and has lower downside risk. Drag the slider to see where the crossover lands for your assumptions.

The market flipped

Six months ago, most market observers expected H100 rental rates to keep falling as Blackwell supply ramped. The opposite happened. H100 1-year contract pricing shot up almost 40%. Firms described trying to find compute in early 2026 as “like trying to book airplane tickets on the last flight out.” ^{[1]SemiAnalysis, "The Great GPU Shortage: Rental Capacity, H100 1 Year Rental Price Index" (April 2026)}

There were several forces that drove the reversal. Demand increased with adoption of compute-heavy agentic AI, while open-source models expanded the customer pool. Supply was also lower than forecasted. Increased memory prices drove up new server costs, stopping potential compute suppliers, and Blackwell procurement lead times extended into mid-2026. ^{[1]SemiAnalysis, "The Great GPU Shortage: Rental Capacity, H100 1 Year Rental Price Index" (April 2026)}

According to SemiAnalysis, all of the Blackwell capacity that they see coming online before September 2026 is already booked. And half the providers in the SemiAnalysis index are completely sold out of the previous generation's Hopper capacity. ^{[1]SemiAnalysis, "The Great GPU Shortage: Rental Capacity, H100 1 Year Rental Price Index" (April 2026)}

For income-based valuation methods like discounted cash-flow, hardware resale values (ie. residual value) are easily derived from the rates the hardware can generate. Reality shows that this is unreliable; the neocloud providers are focused on having stable rates to meet payback requirements. Hardware resale values don't always track the rental rates.

H100 resale prices have climbed with the demand rebound, but eventually, prices should resume declining. That lengthens payback periods for operators and makes the H100 less desirable. A used H100 that you can resell for $15,000 today might only get you $10,000 if you try to resell it in another year. ^{[2]SemiAnalysis, GPU Rental Price Dashboard and ClusterMAX (2025-2026)}

Hyperscalers disagree on how fast to depreciate. Amazon shortened GPU useful life from 6 to 5 years in Q1 2025, a $700M hit. Google extended to 6 years, saving ~$3.4B in 2023. This measure is imperfect of course. Depreciation accounting implies a GPU is worth $0 after its useful life, but that's clearly not true in the real world. It should only be used as a reference. ^{[3]Princeton CITP, "Lifespan of AI Chips: The $300 Billion Question" (October 2025)https://blog.citp.princeton.edu/2025/10/15/lifespan-of-ai-chips-the-300-billion-question/}

57 companies surveyed. Shaded region shows the 3-5yr range used by 80% of filers. Source: 10-K/20-F filings (FY2024-2025). ^{[4]American Compute analysis of 57 public company 10-K/20-F filings (FY2024-2025)}

Option 1: Hold scenario

Two things to consider:

1. The risk of rates falling and residual values also slipping. What gives you confidence in those rates, can you lock in those rates in long-term offtake agreements? Are you sure utilization of H100s will stay high?
2. The opportunity cost of holding. Could you make more money if you switched to newer hardware?

Option 2: Partial refresh scenario

A partial refresh meaning selling some part of the fleet, and upgrading only a part to newer hardware. CoreWeave is the largest neocloud example: 250,000+ GPUs across 32 data centers mixing H100s, H200s, and GB200 systems. ^{[5]CoreWeave S-1 filing and investor relations (2025-2026)https://investors.coreweave.com}

Collateral constraints

If your current hardware was purchased with an active loan the lender has a legal claim on the GPUs as collateral. You cannot sell any of them without lender consent. Even partial sales might break loan covenants, leading to expensive penalties. There's two options:

1. Some lenders will allow a partial collateral release, letting you sell some GPUs, if the remaining loan balance stays below 60% of the value of the hardware you keep. ^{[6]CoreWeave $8.5B financing facility press release (March 2026)https://investors.coreweave.com/news/news-details/2026/CoreWeave-Closes-Landmark-8-5-Billion-Financing-Facility-Achieving-First-Investment-Grade-Rated-GPU-backed-Financing/default.aspx}
2. The more common path is to refinance: use any hardware sale to pay off the existing loan and take out a new loan that covers remaining and new hardware as collateral. Same as the first loan, there'll be one-time origination fees (typically 3% of the new loan amount) plus legal and appraisal expenses.

Not modeled here is that Section 1245 tax recapture is also a cost. If you depreciated the H100s aggressively, their book value in your tax returns were well below what you'll now sell them for, and the IRS taxes that difference as ordinary income. ^{[7]IRS Publication 946, "How To Depreciate Property" and Section 1245 recapture (2025)https://www.irs.gov/publications/p946}

Two things to consider:

1. The risk of losing your market. What gives you confidence that customers will want either generation, and that utilization will stay high? If now you can only offer customers smaller cluster size because you sold some hardware, will that lose you deals?
2. Doing the work, but not realizing the full benefits. You're going through the work and pain of procuring, refinancing, and migrating to a new generation. Sometimes, it's better to whole-heartedly bet on one generation.

Option 3: Full refresh scenario

Two things to consider:

1. Demand is lower than expected. Are you sure new generation's rates and customer demand will be sufficiently high? The hardware is more expensive, so the rates need to be higher to breakeven.
2. Timing is everything. Are you sure you can get the new hardware in time, when procurement is tough? Can you find a facility that supports liquid cooling and the higher power density? Cooling is the colo provider's infrastructure. But available liquid-cooled capacity is scarce, so finding a slot is a search problem. If you're too slow, you might lose customers to competitors.

Migration downtime is not modeled here: procurement, racking, and burn-in for B200s typically takes 1-3 months, and that's a window of zero revenue. At B200 rates that could be $0.5-1.5M in lost income.

How to decide

A few variables determine the path: loan status, current utilization, facility and equipment availability, and contracted demand.

In practice, most operators won't do a clean refresh. Selling hardware that is under contract means downtime, and you can't time it so every long-term agreement expires at once. More likely, you can only sell a portion of the fleet as contracts roll off, and if you pre-sold capacity months ago, even that portion is locked up.

The path of least resistance is to keep the current fleet running until operating costs approach revenue, and deploy new hardware in a separate facility with fresh capital. That is what is happening across the neocloud sector: more debt, more deployments, not refresh cycles. It feeds the growth narrative, but it also means the industry is accumulating hardware rather than replacing it.

Hold when

• Lender prevents hardware sale and refinancing is not available.
• You have high-value, long-term offtake agreements for the H100s.
• No signed or near-signed customer demand at B200 rates.

Partial refresh when

• Customer demand exists at B200 rates, but liquid-cooled colo space is scarce or limited.
• Migration downtime is not acceptable.
• Rubin volume availability is less than 18 months away.

Full refresh when

• Loan clear or refinancing available.
• Customers signing contracts for B200s at $3.50+/hr.
• Liquid-cooled facility secured.
• Downtime during migration is acceptable.
• Rubin volume is 18+ months out.

References

1. SemiAnalysis, "The Great GPU Shortage: Rental Capacity, H100 1 Year Rental Price Index" (April 2026)
2. SemiAnalysis, GPU Rental Price Dashboard and ClusterMAX (2025-2026)
3. Princeton CITP, "Lifespan of AI Chips: The $300 Billion Question" (October 2025)
4. American Compute analysis of 57 public company 10-K/20-F filings (FY2024-2025)
5. CoreWeave S-1 filing and investor relations (2025-2026)
6. CoreWeave $8.5B financing facility press release (March 2026)
7. IRS Publication 946, "How To Depreciate Property" and Section 1245 recapture (2025)