The Non-Depreciation Cycle — When 2020's A100 Started Appreciating and Rewrote the Hyperscaler ROI Model

For three days at the front of the week, one small data point — rental rates for Nvidia A100 GPUs, a card first shipped in Q2 2020, are still rising in August 2026 — has done more to reprice the semiconductor cycle than any of the memory records, foundry expansions, or Beijing fab announcements crowding the same news wire. An Aug 3 read of the GPU-as-a-service market shows A100, H100 and H200 rental rates climbing simultaneously, with B200 units commanding premium spot rates. In every prior silicon cycle — CPUs, storage, networking — the previous-generation part traded at a widening discount from the day the successor shipped. This week, that rule broke.

The break matters because hyperscaler capex mathematics has always rested on it. When Morgan Stanley revised its 2026-2028 cloud infrastructure capex forecast to $1.4T on Aug 3 after hyperscaler earnings, the number was only defensible if a $30k GPU deployed today can be depreciated over three or four years toward a residual worth pennies on the dollar. If the residual instead compounds — if the A100 in your rack five years from now rents for more than it did on the day you bought it — the depreciation charge collapses, ROIC jumps, and $1.4T becomes a floor rather than a ceiling.

The mechanism is memory-shaped, and every other headline on the wire this week fits into it. Conventional DRAM and NAND flash prices set fresh record highs in July — the fourth consecutive monthly record in a shortage that Team Group's CEO called persistent "through 2027 minimum" as the memory-module maker posted a 200-fold Q2 profit surge. Spot DDR5 16Gb cleared $51.33 today, roughly 3.3x the trough set two summers ago. The AI GPU has become memory-bound, not compute-bound, and the resulting inability to feed the next-generation cards at rated bandwidth is what keeps the previous generation economically alive.

Renesas made that arithmetic explicit on Aug 3 by shipping Gen 3 DDR5 MRDIMM parts at 16,000 MT/s — a 60% bandwidth uplift for existing platforms without a socket change. Read the release backwards: the company is selling upgrade parts for servers that would, in a normal cycle, be halfway through retirement. When memory upgrades restore new-generation bandwidth to legacy sockets, the legacy socket depreciates like real estate rather than silicon.

The rating agencies are already re-underwriting the tape. Moody's on Aug 4 handed SK Hynix its first-ever A grade, citing "financial gains from the AI memory boom" — the first time in the memory industry's history that a rating body has treated HBM cash flow as structural rather than cyclical. Barron's simultaneously argued SK Hynix has replaced Nvidia as the single most important AI chip stock, and a separate note laid out an 80% upside case tied to the SanDisk memory-technology partnership announced this week. The theme is monotonic: whoever owns the bandwidth wins the cycle, and the bandwidth constraint is now durable.

That leaves Nvidia in an unusually forgiving spot. Custom-silicon diffusion narratives have been the bear case all summer — MediaTek's $5B ASIC business, Alphabet's Broadcom pivot with its cited $252B unlock, and Google's cited "trailing" position in AI chips. But those bear cases all assume hyperscalers can substitute away from H-series and B-series parts. If instead the installed base of Nvidia GPUs is appreciating — pulling additional rental revenue out of cards already bought and paid for — the substitution math changes. Every H100 that keeps its rental value is a $30k asset that continues to earn against a zero cost basis. That is the definition of an economic moat that widens with age.

The counter-tape is the utility bottleneck. Texas Governor Abbott on Aug 4 ordered PUCT and ERCOT to audit every data center seeking grid connection, and PJM filed a backstop capacity auction plan with FERC on Aug 3 to plug a shortfall that data-center demand is projected to grow through 2030. DTE's $30B capex plan explicitly tied Michigan rate stability to data-center project cadence. The GPU may appreciate; the interconnect queue may not. Any thesis that follows from the rental data has to be underwritten against a power-availability constraint that is now being audited rather than expanded.

Positioning read. The trade this week is not Nvidia versus Micron versus SK Hynix — it is the recognition that AI silicon has left the depreciation curve. For NVDA holders, that reframes forward earnings quality: a rising fraction of the installed base earns rental yield rather than second-hand discount. For MU and the HBM chain (Korea houses the leader, but MU carries the US-listed exposure), the Moody's A-rating template gives the memory oligopoly explicit permission to price for a permanent bandwidth premium. For AVGO and MRVL, the custom-silicon narrative still runs, but its ceiling is now set by how much of the hyperscaler capex tail the appreciating Nvidia installed base absorbs. Watch AMD's print tonight for whether the same rental-market logic extends to MI-series — if it does, the depreciation break stops being a Nvidia story and starts being a category redefinition.

Key Sources: - GPU Rental Market Shows No Depreciation; Rates Rise Across NVIDIA Portfolio (Google News, 2026-08-03) - Renesas Gen 3 DDR5 MRDIMM Targets AI Server Memory Bandwidth Bottleneck (EE Times, 2026-08-03) - Team Group 2Q26 profit surges 200-fold as DRAM shortage persists through 2027 (Google News, 2026-08-04) - Morgan Stanley's $1.4T Cloud Capex Forecast After Hyperscaler Earnings (Google News, 2026-08-03) - Moody's Grants SK Hynix First 'A' Grade Rating, Citing Financial Gains from AI Memory Boom (Google News, 2026-08-04) - plus 76 more