Amazon is accepting billions of dollars’ worth of Nvidia AI systems before every rack has electricity. Microsoft has also acknowledged chips waiting for powered data-center space. Meanwhile, high-memory GeForce cards are scarce and punishingly expensive. The AI buildout is likely contributing to the squeeze on home-AI builders—but the public evidence does not prove Nvidia engineered the shortage or forced customers to buy hardware they did not want.[2][6][8]
The strongest evidence is inside Amazon
Business Insider reported that AWS introduced an “Early Rack Receipt” process under which it expected to receive roughly $5 billion of Nvidia GB300 server racks at US facilities in August, followed by about $3 billion per month “for the foreseeable future.” Some locations had floor space but insufficient power, leaving completed systems in intermediary warehouses. One AWS team had no “clear forecasted path” to install its incoming racks at the time.[2]
Amazon previously took ownership when a rack reached a data center. Under the new process it can pay earlier, before the hardware is producing cloud revenue. Amazon told Business Insider that staging completed equipment near deployment sites is normal and disputed the suggestion that AWS as a whole lacked a plan for those systems.[2]
The problem is broader than Amazon. Microsoft CEO Satya Nadella said in a November 2025 podcast that Microsoft had “a bunch of chips sitting in inventory that I can’t plug in.” He said the constraint had shifted from chip supply to powered, ready-to-use data-center space. Morgan Stanley later estimated that US data-center developers face a 34% net power shortfall through 2028, equivalent to 32 gigawatts.[4][8]
The warehouses and power bottleneck are documented. The harder question is why buyers accept expensive equipment months before all of it can earn money.
Where the channel-stuffing theory begins
Writer and analyst Ed Zitron argues that Nvidia uses scarcity to encourage early acceptance: buy now or risk losing an allocation when a data center is finally ready. He calls this “deliberate channel stuffing.” RespawnFirst extended the argument to gaming, alleging that Nvidia routed manufacturing capacity away from GeForce cards and into AI hardware that can sit unpowered.[1][3]
That is a serious accusation, not a synonym for forward buying. Channel stuffing implies that shipments are pulled ahead of genuine demand to inflate current sales, often by pushing excess inventory into a distribution channel. The reporting here shows early acceptance and delayed deployment. It does not establish excess unwanted inventory, special incentives, unusual return rights, improper revenue recognition or contractual coercion. TechRadar’s review reached a similarly cautious conclusion: Nvidia’s AI priority is visible, but channel stuffing remains an interpretation layered onto the underlying reporting.[7]
There is still a credible consumer-supply story. Tom’s Hardware reported an unverified claim that Nvidia stopped supplying the GB202 processor for the GeForce RTX 5090 so the same silicon could be used in data-center and professional products. The RTX Pro 5500 has the same 21,760 CUDA-core count as the RTX 5090 but carries 84GB of memory and sells into a much richer market. If that report is accurate, AI and professional demand is directly competing with Nvidia’s highest-memory consumer card.[5]
The overlap is real but easy to exaggerate. Consumer Blackwell chips use TSMC’s 4N process, while the B200 data-center chip uses the related 4NP process. Complete AI systems also require different memory, packaging, boards, networking and power hardware. A GB300 rack sitting in a warehouse cannot simply be turned into retail 5090s. The clearest competition is between products using the same large GB202 die.[5][9]
Nvidia is expanding the supply chain around AI. On October 21 it announced agreements covering substrates, advanced packaging, testing and optical engines. The company said its partners would support a 2.5-fold increase in substrate capacity over three years, while Amkor aims to increase advanced-packaging capacity by more than 300% by the end of 2028. That demonstrates an AI-first expansion strategy; it does not prove fraudulent or coercive selling.[10]
The consumer damage is no longer theoretical
Tom’s Hardware’s live US tracker recently showed an RTX 5090 marketplace listing at $8,599 versus a $1,999 launch price, while first-party 5090 stock was effectively absent. That is an asking price from a third-party seller, not a representative completed sale. The same tracker showed an RTX 5080 listing at $1,899 versus $999 and a 16GB RTX 5060 Ti at $765 versus $429, and attributed broader price pressure partly to enormous memory orders for AI data centers.[6]
For a home-AI builder, the RTX 5090’s 32GB of VRAM is the attraction. A 30-billion-parameter model quantized to four bits occupies about 14GiB in raw weights, leaving useful room for context and runtime overhead. A 70B model is roughly 32.6GiB before overhead, so it will not fit completely on a single 5090. Paying a panic premium still leaves you short of that next useful model class.[6][9]
A used 24GB RTX 3090 or 4090, an Apple Silicon system with ample unified memory, or a carefully planned multi-GPU build may offer better value. But two GPUs do not automatically behave like one larger card: the inference runtime must support model splitting, and the builder must account for motherboard lanes, power, cooling and communication overhead. Apple reserves part of unified memory for macOS. AMD and Intel cards can run smaller local models, but buyers should verify that their exact application supports the required ROCm, Vulkan, Metal or oneAPI backend before purchasing.
What the evidence actually supports
We can say three things with confidence. Hyperscalers are receiving AI hardware faster than every system can be powered. Nvidia and its customers are securing scarce supply despite that infrastructure mismatch. Reported prioritization of AI and professional products, combined with AI-driven demand for memory, is consistent with—and likely contributes to—the miserable market for high-memory consumer GPUs.[2][4][5]
We cannot yet say Nvidia forced Amazon to buy unwanted racks, deliberately created the GeForce shortage, or diverted a known number of consumer chips into idle AI systems. RespawnFirst and TechRadar are not independent confirmation of that stronger thesis; both build on the same Business Insider reporting and Zitron’s analysis.[1][2][7]
The defensible conclusion is less cinematic than “Nvidia manufactured the crisis.” Nvidia has created an allocation environment in which its richest customers can accept hardware early, AI systems can wait for power, and home users shop from whatever high-VRAM supply remains. That is aggressive prioritization with real consumer consequences. Calling it channel stuffing requires evidence we have not seen.
Don’t buy a 5090 at today’s panic prices. To upgrade this story from aggressive allocation to channel stuffing, we would need an internal policy, customer contract or shipment record showing that early acceptance was a condition of future supply—or financial evidence that Nvidia pushed more inventory than customers genuinely wanted. Until then, the warehouses are documented. The coercion is not.[2][6][7]
Sources
- [1] RespawnFirst: Multiple Reports Say Nvidia Manufactured the Gaming GPU Crisis
- [2] Business Insider: Amazon Buys Billions in Nvidia AI GPU Racks, Then Waits for Power
- [3] Ed Zitron: AI Had Better Work Out
- [4] Reuters: AI power crunch hits chip supply chain
- [5] Tom’s Hardware: Nvidia reportedly halts RTX 5090 production
- [6] Tom’s Hardware: GPU price tracking 2026
- [7] TechRadar: Is the GPU stock crisis down to real shortages?
- [8] TechRadar: Satya Nadella says Microsoft cannot power all its AI chips
- [9] Tom’s Hardware: Nvidia Blackwell and RTX 50-series
- [10] Nvidia: Supply chain partners expand AI capacity