The Eight-Year GPU — The Week CoreWeave's 2029 A100 Contract Rewrote the Depreciation Model

Three events in three days

The second week of August 2026 is when the market silently voted on how long a GPU can keep printing rent. Three events collided inside 72 hours.

First, CoreWeave disclosed a multi-year contract renting its A100 GPUs through 2029. That is a pre-Hopper part, first shipped May 2020, still billing customers into its ninth year. It breaks the 3-year depreciation model hyperscalers use for planning and the 4-year GAAP schedule accountants prefer.

Second, the same company raised GPU hourly pricing by 25%. Q2 revenue beat consensus and the stock ripped 14% after-hours. So the "old" A100 sells not only longer, but at higher rates. Both the demand curve and the residual-life curve moved right at the same time.

Third, Nvidia formalized a program guaranteeing up to 25% GPU residual value as part of the $500B financing package, with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR on the paper. Nvidia is now underwriting — with its own balance sheet — that a GPU still holds a quarter of its cost after five years.

One cell — useful life

All three events collapse into a single spreadsheet cell: GPU useful life. Analysts and hedge managers spent the last three years typing 3 into that cell, generously 4. This week the market quietly typed 6, 7, maybe 8.

The numbers are punishing. One Nvidia Kyber rack carries 340.4TB of DRAM plus HBM4E at $19.76/GB and lists at $41.6M. Pushing useful life from 4 to 6 years drops hourly depreciation per rack by ~33%. That's several hundred basis points of hyperscaler gross margin — which is exactly why CoreWeave still fills bookings after a 25% price hike.

Standing against that logic is Michael Burry. This week he argued Nvidia's depreciation schedule is too slow and that normalized economics would strip 20–30% off hyperscaler EPS. At the moment the market wants to move the cell to 8, he is trying to pull it back to 3.

The critical point: this week's data refuted Burry. A100 rentals are being renewed in year six, at prices 25% higher than before. CoreWeave's Q2 print and the 14% after-hours move are the empirical answer.

The collateral question — $500B's real risk

In Nvidia's $500B financing frame, GPUs are collateral. Collateral value is set by residual value. And this is where the week detonated most quietly.

The threat: China's AI chips. Tech Times reported that China's AI-chip boom threatens the GPU collateral value backing the $500B deal. If a Chinese H100-equivalent lands at half price in five years, Nvidia's 25% guarantee becomes out-of-the-money option writing. A US House member calling this week for enforcement of Biden-era chip customer due diligence rules sits precisely on top of this pressure.

Geographic reassembly is running in parallel. SK Hynix moving to sell its southwest China fab, Amkor weighing a $1B+ stake sale in its China unit, TSMC announcing its largest-ever foreign investment in the US — same week, same logic. The industry is trying to detach GPU residual value from origin risk.

What the eight-year story leans on

The memory market backs the thesis. DDR5 16Gb spot printed $52.5 on Aug 12, driving 8%/8%/4% jumps in SK Hynix, SanDisk, and Western Digital. Apple cut 2026 hardware shipments as DRAM prices rose 29%. That's exactly why CoreWeave is clinging to A100s — there isn't enough HBM/DRAM to stand up replacement racks at will.

Supply also defends the eight-year story. Foxconn confirmed Q4 Vera Rubin shipments start this quarter, but the $41.6M/rack sticker makes throwing out A100s economically absurd. The classic semiconductor cycle rule — new gen arrives, old gen goes to scrap — is not holding this cycle.

PM note

The real Nvidia long thesis is no longer revenue growth. It is whether the useful-life cell stays at 6–8. This week CoreWeave, Nvidia, and the market wrote 8 into it. Burry is trying to defend 3.

The gap between those two numbers is a 20–30% EPS swing across the hyperscaler complex and, ultimately, the strike-adjusted value of Nvidia's residual guarantee. When and how the $500B collateral pool gets remarked is the only real Nvidia risk question for the next two years.

On the memory side, the same story translates into a cleaner long. SK Hynix confirmed a $38B HBM capex expansion; Mizuho holds a $1,375 target on Micron. A rack that lives eight years doesn't refresh HBM/DRAM every five — it consumes memory across its entire life. That reshapes the total memory demand curve upward, not just its slope.

Bottom line: this week the market consensus on "how many years an AI rack earns" migrated from four to six-to-eight. Nvidia put its own balance sheet behind a 25% floor on that migration. While the consensus holds, $500B is a residual-value call option, not circular financing. The moment it cracks, it becomes a collateral event.

Key Sources: - CoreWeave's 2029 A100 contract challenges one-term GPU depreciation model (Blockspace Media, 2026-08-12) - Nvidia Guarantees 25% GPU Residual Value to Unlock $500B AI Financing (eGamers.io, 2026-08-11) - CoreWeave Raises Prices 25% Amid Strong AI Demand (Benzinga, 2026-08-12) - Nvidia and Burry clash over GPU depreciation amid $500B AI financing push (Pluang, 2026-08-11) - NVIDIA Kyber Racks Configured With 340.4TB DRAM, HBM4E At $19.76/GB (Wccftech, 2026-08-11) - plus 55 more