How Long Can an AI Chip Earn?
Separating prompt processing from answer generation can give older GPUs another job later in life. The reviewed loans still depend on the original customer's payments.
Read the companion deck (PDF)→
Prompt processing and answer generation can run on different hardware
Investor Gavin Baker suggested in a May 2026 podcast that separating the two jobs involved in AI inference could keep NVIDIA's older Ampere or Hopper GPUs productive alongside Cerebras or Groq hardware for 10 to 15 years. Longer earning lives would matter to borrowers and lenders only if the machines kept producing revenue after today's customer contracts expire. We think disaggregated inference can preserve useful work for older GPUs. We would not treat years 6 to 10 as dependable base-case cash flow.
An AI model first processes the prompt and builds a temporary working memory known as the KV cache. It then generates the answer one token at a time while consulting that cache. Engineers call the two phases prefill and decode. Prefill wants a lot of computation quickly, while decode often waits on memory and communication. Running them on different machines can match each job to the hardware that handles it best, but the cache has to move fast enough to avoid adding more delay than the split removes.
On February 18, 2026, NVIDIA and Sarvam AI reported that an H100 system serving one model with 30 billion learned weights raised its completed-token throughput index from a 1.00 baseline to 1.26 with faster chip-level software, 1.31 with better request scheduling, and 2.00 after the two phases were separated and tuned. The system maintained 75 output tokens per second per user, and the split itself accounted for roughly a 1.5 times step. This was a vendor-authored demonstration with one mix of prompt and answer lengths, one model, and H100s on both sides of the split.
AWS's SageMaker HyperPod work, reported July 10, 2026, found the strongest gains with long prompts, many simultaneous users, and answers delivered as they were generated. For requests processed in groups, lighter traffic, and short prompts, keeping both phases together was simpler. AWS set 4,096 input tokens as the default routing threshold. Below that level, moving the KV cache cost more time than separating the two phases saved.
HMA-Serve's June 2026 research prototype moved the cache over a 100-gigabit-per-second direct-memory network between Tenstorrent hardware doing prefill and NVIDIA A100s doing decode. It reported up to 3.2 times more useful completed throughput and 4.8 times more useful throughput per dollar. The A100's decode role shows that the architecture does not permanently assign an older GPU to prefill, where Baker's suggestion places it. Two 2026 announcements describe future production deployments. AWS and Cerebras would connect Trainium prefill to Cerebras decode for Bedrock, while NVIDIA's Rubin and Groq design would require more networking, control hardware, and liquid cooling. HMA-Serve covered four Qwen3 language models and three recorded workload patterns; the announced deployments use current-generation components and supply no later-life production economics (Exhibit 1).
EXHIBIT 1
The split works on H100s; an old-vintage fleet remains unobserved
Disaggregated-inference evidence by maturity stage: who runs prefill, who runs decode, and what links them
Scheduler and router allocate requests across both pools
| Evidence | Prefill | Decode | Transfer & system layer | What it proves |
|---|---|---|---|---|
| NVIDIA + Sarvam AI | Prefill: H100 | Decode: H100 | SGLang router, 1P+1D KV transfer | What it proves: Kernels, scheduling, and the split lifted one H100 case to 2.0x; the split alone about 1.5x |
| AWS + Cerebras | Prefill: AWS Trainium | Decode: Cerebras CS-3 | Amazon EFA; Bedrock, forthcoming | What it proves: Commercial design assigns phases to different accelerator types |
| NVIDIA Rubin + Groq 3 LPX | Prefill: Rubin GPU | Decode: Groq 3 LPU | Fabric logic, BlueField-4, liquid-cooled rack | What it proves: Current product pairs GPUs, LPUs, network, and rack |
| HMA-Serve (preprint) | Prefill: Tenstorrent Blackhole | Decode: A100 80GB (old) | RDMA, cross-vendor KV-format conversion | What it proves: Prototype puts an A100 80GB on decode; higher goodput per dollar |
| Life-extension claim | Prefill: Hopper or Ampere | Decode: Cerebras or Groq LPU | Required, not quantified | What it proves: Proposed 10 to 15 year vintage extension; no operator result |
Note: NVIDIA and Sarvam use H100 for both phases; the split contributes roughly 1.5 times for the tested model, workload and latency target. AWS-Cerebras and NVIDIA Rubin-Groq use current-generation components. HMA-Serve assigns an A100 80GB to decode. No case reports post-year-five operator economics.
Source: NVIDIA, Cerebras, AWS, HMA-Serve and Gavin Baker.
Keeping H100 servers versus buying B200 servers
At a power-bound site, adding one server requires removing or displacing another, so output per kW and rack matters. The MLCommons v5.0 completed-token results and NVIDIA power specifications describe complete DGX servers, each containing several GPUs. They put a DGX H100 server at 31,106 tokens per second using up to 10.2 kW and the newer DGX B200 server at 98,443 using up to 14.3 kW. On that basis, the B200 server produces about 2.3 times as many completed tokens for each maximum input kW.
MLCommons measures completed tokens per maximum input kW. The NVIDIA and Sarvam AI single-model site cases in Exhibit 2 use different service speeds, rack layouts, and denominators to answer different operator questions. Across them, the B200 advantage ranges from parity to 2.8 times as much throughput. Exhibit 2 carries the keep-versus-replace crossover at these assumptions.
An already owned H100 still incurs power and other avoidable operating costs. Our model assumes six cents of other avoidable operating cost for every kWh of electricity used by the system. At a site with no better use for its power and rack space, electricity prices of 4, 8, and 12 cents per kWh put the simple cash costs of keeping the H100 at about $0.009, $0.013, and $0.016 per million tokens.
The B200 must also repay its purchase price, and it can do so only through added tokens, the output beyond what the displaced H100 was already producing. Using $32,000 per nameplate kW, spreading the purchase across the added output requires roughly $0.098 to $0.265 per million added tokens over two to three years at 50 to 90 percent utilization. At three years and 70 percent utilization, the required price is about $0.126 per million tokens.
DeepSeek's first-party hosted API lists V4 Flash output at $0.28 per million tokens, about 2.2 times the $0.126 base crossover. The June 2026 Silicon Data token index of publicly listed model prices showed about $1.75. The $0.28 figure is one public customer-price anchor; self-hosted costs and negotiated contracts can differ. The calculation also omits the specialized processor paired with the older GPU, network and integration work, downtime, the H100's sale value, different utilization for old and new servers, and the cost of capital.
At a power-bound site, the simple model generally favors replacing H100 servers with B200 servers. Retention can still make sense when routing software sends suitable work to a split system, fast cache transfer gives an A100 a useful role in a different phase, or an older high-memory fleet serves a cheaper or otherwise differentiated workload. Those cases become more plausible with cheap power, spare electrical and rack capacity, or low site opportunity cost.
The disclosed GPU debt clocks in this sample end by year six
Exhibit 3 compares three clocks: management's equipment-life estimate, the customer's term and debt maturity. A second workload may extend useful life; financing it still requires customer cash flow during the loan. If the customer leaves before the asset or debt ends, the operator must find another buyer while principal is still due. Deployment delays, contract termination, operating costs, customer concentration and resale risk can all break the match.
Five public operators report equipment lives of five to six years and named or weighted customer terms of three to five years. Three of the five match exactly; debt is outside that count. IREN is the sole operator with an all-three-clock match tied to the same project. Across the broader financing source set, disclosed GPU debt runs two to six years. Several large private facilities omit maturity, and none of the maturities reviewed reaches the proposed ten-year life-extension horizon.
Nebius and WhiteFiber illustrate the exact asset and customer matches. Nebius moved server life to five years in the first quarter of 2026, citing utilization commitments, and its named Microsoft and Meta agreements each run five years. WhiteFiber reports a five-year cloud-equipment life and five-year Paris contract. Its two-year Iceland GPU loan can extend to four years, but belongs to another fleet.
Applied Digital Cloud and CoreWeave supply company-wide figures without a project-level match. Applied Digital Cloud uses five-year amortization for most equipment finance leases against a 36-month named second AI-customer agreement; filings map no lease to that customer. CoreWeave's six-year technology-equipment life and roughly five-year weighted committed-contract duration at year-end 2025 map no customer to a particular GPU pool.
IREN reports a five-year average for both its high-performance computing hardware life and Microsoft GB300 service agreement. Approximately $3.6 billion of project financing is due by December 31, 2031, or earlier when Microsoft pays the last service fee. Repayment occurs during the contract, and lenders take security over the project GPUs, project equity, and contract cash flow. IREN has not disclosed the interim repayment schedule; the latest possible repayment date cannot extend beyond the final disclosed Microsoft payment.
Management has extended equipment lives to five or six years
CoreWeave, Nebius and WhiteFiber have all lengthened their equipment-life estimates. A longer estimate spreads an asset's cost across more reporting periods, lowering current depreciation expense and raising reported income. The revisions do not forecast an eventual sale price or convey a lender judgment. The hollow circles in Exhibit 3 mark the prior equipment lives behind these revisions.
CoreWeave extended computing equipment from five to six years effective in 2023, reducing that year's total expense by $20 million, according to its 2025 Form 10-K. Nebius moved servers from four to five years in the first quarter of 2026 based on utilization commitments. Its first-quarter 2026 Form 6-K, Exhibit 99.2, reports $43.1 million less quarterly depreciation and $41.6 million more net income. WhiteFiber moved its disclosed equipment life from three years to five. All three now use five or six years; their stated reasons included usage and customer commitments.
CoreWeave said on its May 2026 first-quarter earnings call that Blackwell, Hopper, and Ampere were all earning revenue; prices rose from the prior quarter for A100, H100, H200, and L40 capacity, and near-term supply was largely sold out. IREN's fiscal 2025 Form 10-K and subsequent Form 8-K show H100 and H200 fleets operating while B200, B300, and GB300 systems are procured, with the new GB300 deployment tied to Microsoft and new project financing. Applied Digital Cloud reported all 6,144 deployed GPUs earning revenue at year-end 2025.
WhiteFiber sold H200 servers in the first quarter of 2026 for $26.1 million against a $24.3 million book value, recording a $1.8 million gain. The filing omits unit age, count, buyer, and reason for sale, so this remains one transaction during a GPU shortage. Around year five, management would still have to choose the old GPU's next workload and location, pay to connect it to any mixed-chip system, secure or renew customer demand, and decide whether its financing continues.
A second workload must cover the old GPU's costs
Across 63 providers, the AIMultiple Cloud GPU Rental Price Index median public on-demand A100 rate fell 41.7 percent from July 2024 to June 2026, from $3.07 to $1.79 per GPU-hour. H100 was roughly flat over the same period; H200 and B200 rose from later starts. A falling old-card rate pushes the operator toward a newer GPU that produces more output from the same electricity and data-center space.
AIMultiple's index consists of asking prices. It does not report negotiated contracts or realized margins, and its provider mix changed during a GPU shortage. Applied Digital Cloud's unaudited sale forecast assumed roughly 40 percent lower hourly prices for next-generation equipment after its contract term. The 40 percent forecast is one sale-model assumption, and the A100 decline is one public-rate path. Both inform the revenue side of the decision; neither establishes how long the equipment remained in service.
If disaggregated inference gives an old card a second workload, its lower revenue still has to cover electricity, hosting, maintenance, integration and the opportunity cost of powered space. Fewer cards would retire as new generations arrive only where that lower revenue covers the added costs.
As new deployments ease a shortage, the newest GPUs would tend to occupy facilities that can supply their electricity and cooling needs. Older high-memory fleets could move to cheaper power, less urgent inference work, or older data centers unable to support the newest systems. Used aircraft, servers, and mining rigs offer examples of equipment moving to a second use after its price falls.
At the previous price, the owner of the older equipment keeps the extra earning years. Falling prices give customers more of the benefit through cheaper compute. Operators would bear the cost and execution risk of integrating mixed-chip systems.
What today's lenders protect
CoreWeave's two 2026 delayed-draw term loans use contracted customer payments to repay the debt. DDTL 4.0 paired one stronger customer with limited parent support, so the lender had less access to CoreWeave if project cash fell short. DDTL 5.0 paired two weaker customers with a full parent guarantee, a 35% cash-flow cushion and an interest spread twice the first loan's; the guarantee put more of any shortfall on CoreWeave.
In March 2026, the $8.5 billion DDTL 4.0 financed one investment-grade customer, unnamed in the public documents and reported by Bloomberg to be Meta, and received A3/A(low) ratings. Its floating interest rate was SOFR plus 225 basis points. The disclosed initial fixed-equivalent borrowing cost, reflecting interest-rate protection, was roughly 5.9%. Contracted cash flow had to equal at least 1.15 times scheduled debt payments, a 15% cushion. The loan was privately placed with Blackstone reported as the anchor investor, and it matures in March 2032.
In May 2026, the $3.1 billion DDTL 5.0 served two below-investment-grade customers and received Ba2/BB+ ratings. It came with SOFR plus 450 basis points, a full CoreWeave parent guarantee and contracted cash flow of 1.35 times scheduled debt payments. It matures in November 2031. Public syndication lowered the final margin over SOFR by 50 basis points. The 225-basis-point difference between the two loans cannot be credited to GPU residual value because customer quality, parent support, ratings, loan rules, repayment schedules and distribution methods all differed.
Both agreements also assign the GPUs a "GPU Depreciated Amount" that falls evenly over six years. That contractual convention is not an appraisal, resale quote or expected recovery. DDTL 4.0 directs customer receipts first to operating costs, loan administration and hedging, then to interest and principal. Reserves and cash traps come next, ahead of any owner distribution.
DDTL 4.0's loan rules identify hardware by serial number and protect warranties, power arrangements, insurance and access to the equipment. Interest-rate hedges limit the effect of a rise in the floating base rate. IREN's Microsoft project gives lenders the same types of claims over the project GPUs, ownership interest and Microsoft contract cash flow, with debt due no later than the final Microsoft service fee.
Specialist lessors would have to take value risk across customers, and independent appraisers would need a recognized way to inspect, certify and value older systems. Another lender would have to finance a specialized processor or network upgrade, or the cost of connecting or moving the equipment midway through its life. Investors would need observed prices for years 6 to 10. Financing would also have to span several projects and customers, with the securities issued against that pool receiving a credit rating.
Aircraft finance assembled the same set of institutions, and the assembly took two decades. ILFC, the International Lease Finance Corporation, was founded in 1973 to buy aircraft and lease them to airlines. Aircraft age through flight cycles and regulated maintenance; GPU generations face rapid technological change. US Section 1110 strengthened repossession rights in 1978. ISTAT, founded in 1983, later established standards for independent appraisers. GPA's 1992 ALPS securitizations bundled aircraft and lease claims for sale to investors, and the first enhanced equipment trust certificate in 1994 added another structure for issuing rated securities against a pool of aircraft.
EXHIBIT 5
Aircraft credit built its residual institutions over two decades
Three financing layers, the aircraft market against the GPU market in 2026
| Financing layer | Aircraft market | GPU market, 2026 | Status |
|---|---|---|---|
| Contract-backed senior debt | Established: equipment trusts, Section 1110, EETCs from 1994 | Observed: IREN to 2031; CoreWeave DDTL 4.0 and 5.0 | Observed |
| Residual owner, appraisal, midlife reserve | Established: lessors from 1973, ISTAT appraisers from 1983, maintenance reserves | One fund-lease vehicle disclosed (Valor-xAI); no appraiser, no reserve | Partial |
| Diversified securitization of residual value | Established: pooled lease deals from 1992 | DDTL 5.0 is one borrower syndicated; no rated GPU pool | Absent |
Aircraft became financeable beyond the first contract only after lessors, appraisers, maintenance reserves, and pooled deals made the residual observable. GPU credit has built the senior layer and none of the rest.
Note: Aircraft age by cycles under regulated maintenance; GPUs age by technology. The comparison concerns the institutional sequence. Valor-xAI's lease of GB200 systems to an xAI subsidiary is the nearest current GPU example, with no disclosed later-life residual.
Source: ILFC and GPA histories; ISTAT; equipment-trust primers; company filings; Second Order framework.
Three events would show post-contract financing works
Can a named GPU generation still generate enough cash or sale value after roughly year five to support a new loan? Answering that question requires disclosed post-contract economics for the fleet.
We would need its assigned job, utilization, realized revenue and margin, and power and hosting cost. We would also need the spending on new accelerators, networks or site work required to keep it useful, plus the customer's term and renewal, the loan's repayment schedule and maturity, and the hardware's eventual sale or recovery value.
Three events could provide a clean test. An operator could publish economics for a specific vintage after year five. A third party could buy a fleet and disclose its age, unit count and price. Or a lender could advance money after the original contract ends, relying on the old fleet's later cash flow or recoverable sale value.
An AI model does two different jobs when it answers a question. It first processes the prompt and builds a working memory, then produces the answer one small unit of text at a time. The first job wants a burst of computation; the second often waits on memory and communication. Running them on different machines can give an older GPU a useful role alongside a newer or specialized processor, provided the working memory moves between them quickly enough.
NVIDIA and Sarvam AI reported in February 2026 that separation and related tuning lifted completed-token throughput on one H100 system from a 1.00 baseline to 2.00 while maintaining 75 output tokens per second per user. The split itself accounted for roughly a 1.5 times step. A June 2026 research prototype paired Tenstorrent hardware with NVIDIA A100s and reported as much as 3.2 times more useful throughput and 4.8 times more useful throughput per dollar. AWS found in July 2026 that the split worked best for long prompts and heavy simultaneous use. Below its 4,096-token routing threshold, moving the working memory cost more time than the separation saved.
The economic hurdle rises when electricity and rack space are already full. MLCommons v5.0 results and NVIDIA specifications imply that a complete B200 server produces about 2.3 times as many completed tokens per maximum input kW as an H100 server. A purchased B200 repays itself only through added tokens, the output beyond what the displaced H100 was already producing. Our model spreads the purchase price, at $32,000 per nameplate kW, across three years of that added output at 70% utilization. The result is a required selling price of about $0.126 per million added tokens. DeepSeek lists V4 Flash output at $0.28, a single first-party customer-price anchor, about 2.2 times the requirement. Realized operator margin can differ from that public price. The comparison also excludes integration work, downtime, financing cost and the H100's sale value, so it does not establish payback. Cheap power, spare capacity or a differentiated workload can still make retention sensible.
Five operators report equipment-life estimates of five to six years and customer terms of three to five years; a broader financing sample puts disclosed GPU debt at two to six years. IREN is the one operator in the sample that aligns the equipment life, Microsoft contract and project debt on the same five-year project. The debt is due no later than Microsoft's final service payment. CoreWeave's 2026 loans likewise rely on contracted cash flow, guarantees, reserves and control of identified GPUs. Their six-year contractual depreciation convention supplies no market appraisal.
NVIDIA has demonstrated the split on H100s, AWS and Cerebras have announced a production design for Bedrock, the NVIDIA Rubin and Groq design is announced, and the HMA-Serve prototype already pairs Tenstorrent hardware with NVIDIA A100s. If disaggregated inference delivers later-life economics, we would expect GPU financing to adapt in the order aircraft finance did: lenders advancing more against the same machines and letting repayment run past the first contract, specialist owners taking value risk across customers, the role ILFC filled from 1973, then independent appraisers and eventually pooled, rated financings. Aircraft finance took two decades to complete that build, and GPU generations turn over faster than airframes, which could compress the sequence or cut it short. Retirements would slow as each new generation entered service, and the split of the gains between owners and customers would depend on how far compute prices fall as supply expands. Our base case assigns no earnings to years 6 to 10 until one of three events appears in a filing: an operator reporting post-contract economics, a sale that prices a used fleet, or a loan that advances against later-life cash flow.
- Gavin Baker, Invest Like the Best, episode 473, "Watts and Wafers" (May 20, 2026), the life-extension claim this issue tests: https://podscripts.co/podcasts/invest-like-the-best-with-patrick-oshaughnessy/gavin-baker-watts-and-wafers-invest-like-the-best-ep473
- NVIDIA and Sarvam AI, "How NVIDIA Extreme Hardware-Software Co-Design Delivered a Large Inference Boost for Sarvam AI's Sovereign Models" (February 18, 2026): https://developer.nvidia.com/blog/how-nvidia-extreme-hardware-software-co-design-delivered-a-large-inference-boost-for-sarvam-ais-sovereign-models/
- AWS Machine Learning Blog, "Disaggregated prefill and decode for LLM inference on SageMaker HyperPod" (July 10, 2026): https://aws.amazon.com/blogs/machine-learning/disaggregated-prefill-and-decode-for-llm-inference-on-sagemaker-hyperpod/
- HMA-Serve research preprint, arXiv:2606.29986 (June 2026): https://arxiv.org/abs/2606.29986
- MLCommons Inference v5.0 completed-token results and NVIDIA system specifications: https://docs.mlcommons.org/inference_results_v5.0/ ; https://docs.nvidia.com/dgx/dgxh100-user-guide/introduction-to-dgxh100.html ; https://www.nvidia.com/en-us/data-center/dgx-b200/
- Realized-price anchors: DeepSeek first-party API pricing, V4 Flash output at $0.28 per million tokens: https://api-docs.deepseek.com/quick_start/pricing ; Silicon Data LLM Token Expenditure Index (SDLLMTK), June 2026 print of about $1.75 per million tokens: https://portal.silicondata.com/token-index-chart
- AIMultiple Cloud GPU Rental Price Index, public on-demand rates across 63 providers (retrieved July 12, 2026): https://aimultiple.com/gpu-index
- CoreWeave 2025 Form 10-K and first-quarter 2026 earnings-call materials: https://www.sec.gov/Archives/edgar/data/1769628/000176962826000104/crwv-20251231.htm ; https://s205.q4cdn.com/133937190/files/doc_financials/2026/q1/CoreWeave-Inc-CRWV-US-Q1-2026-Earnings-Call-7-May-2026-5_00-PM-ET.pdf
- CoreWeave DDTL 4.0 and DDTL 5.0 credit agreements: https://www.sec.gov/Archives/edgar/data/1769628/000176962826000129/ex1011.htm ; https://www.sec.gov/Archives/edgar/data/1769628/000176962826000236/ex101.htm
- Bloomberg, "CoreWeave Raises $8.5 Billion GPU Loan Backed by Meta Deal" (March 31, 2026): https://www.bloomberg.com/news/articles/2026-03-31/coreweave-crwv-raises-8-5-billion-gpu-loan-backed-by-meta-deal
- Nebius 2025 Form 20-F and first-quarter 2026 Form 6-K: https://www.sec.gov/Archives/edgar/data/1513845/000110465926052948/nbis-20251231x20f.htm ; https://www.sec.gov/Archives/edgar/data/1513845/000110465926064092/nbis-20260331xex99d1.htm
- IREN fiscal 2025 Form 10-K and Microsoft project financing Form 8-K: https://www.sec.gov/Archives/edgar/data/1878848/000187884825000063/iren-20250630.htm ; https://www.sec.gov/Archives/edgar/data/1878848/000114036126023427/ef20075181_8k.htm
- WhiteFiber 2025 Form 10-K and first-quarter 2026 Form 10-Q: https://www.sec.gov/Archives/edgar/data/2042022/000121390026034341/ea0278305-10k_white.htm ; https://www.sec.gov/Archives/edgar/data/2042022/000121390026056212/ea0288695-10q_whitefiber.htm
- Applied Digital Cloud transaction information statement: https://www.sec.gov/Archives/edgar/data/1549084/000143774926010424/ekso-20251231.htm
- Aircraft-finance institutional sequence in Exhibit 5: ILFC and GPA company histories; ISTAT Certified Appraiser program materials; Section 1110 and equipment-trust-certificate primers; GPA's 1992 ALPS lease-portfolio securitizations.
This issue was edited for readability after first publication. The conclusion, data, and sources are unchanged.
Second Order is an independent research briefing on AI's macro impact. It is provided for discussion purposes only and is not investment, legal, tax, or accounting advice. © 2026.