The GPU Shortage Is Not Over. It's Just Different.

Nvidia's CFO confirmed RTX 5090 supply constraints will persist through 2027. That sounds like a supply story. It is not. It's a market fragmentation story.

Ask most people whether the GPU shortage is over and they'll say yes. Cloud H100 rents are cheaper. Imported 4090s are easier to find. Some AI Twitter account declared victory. But dig into what kind of GPU work is actually happening, and a more complicated picture emerges. The shortage never ended. It just moved.

Where the Scarcity Actually Lives

The RTX 5090 launched with a $1,999 MSRP. It sells for $2,800 to $3,500 at retail, when it is in stock at all. Nvidia's own CFO, Colette Kress, confirmed supply constraints will continue through early 2027 during an earnings call. That is not a rumor. That is a public statement from the company's finance chief.

The demand is not abstract. Gaming demand from Southeast Asian markets has grown substantially. Ethereum mining has revived on weaker coins. AI developers compete with datacenters for the same silicon bin. The result is a consumer GPU market that remains structurally tight, regardless of what cloud rental rates suggest.

This matters because a large share of the "AI inference is getting cheap" narrative is built on cloud H100 pricing, not local hardware ownership. Owning a 5090 for personal AI work is still a secondary market ordeal.

The Card That Changed the Local AI Story

Something interesting happened in the workstation segment. The NVIDIA RTX PRO 5000 72GB launched in December 2025 and reached general availability in early 2026. It is not a consumer GPU. It does not play Cyberpunk at 4K. It is a workstation part priced for serious AI workloads.

The 72 gigabytes of GDDR7 memory is the spec that matters. Most consumer cards top out at 24GB. That ceiling prevents serious local inference on the models that developers actually want to run. A 70-billion parameter model at 4-bit quantization needs roughly 40GB just to load. On a 24GB 4090, you are swapping to CPU RAM constantly or running a quantized version that loses performance. On the RTX PRO 5000, you load the model and stay in VRAM.

The "Shortage Is Over" Narrative Is Wrong and Why

The misunderstanding comes from conflating different GPU markets with different demand drivers.

Cloud GPU availability has genuinely improved. Rental rates for H100 instances have dropped from 2024 peaks. This reflects datacenter supply catching up to enterprise AI demand, not consumer GPU markets normalizing.

Consumer markets tell a different story. The RTX 4090, a card that launched in late 2022, still commands premium pricing in 2026 because demand from AI developers, gamers, and miners continues to exceed supply at the original MSRP. The RTX 5090 shortage compounds this by pushing 4090 demand upward as a fallback option.

The cloud compute advantage is real but bounded. Running inference on a rented H100 at $2 to $3 per hour is feasible for development and testing. It is not feasible for production workloads at scale, or for developers who need to run models continuously without watching a meter run. Local inference has a permanent cost advantage in long-running scenarios, which is why serious AI developers want GPU ownership, not just cloud access.

The Fragmentation Is the Story

Here is what the GPU market actually looks like in 2026: a tight consumer segment where the 5090 shortage keeps 4090 prices elevated and MSRP cards are effectively unavailable; a professional segment where the RTX PRO 5000 72GB has created a new tier of desktop AI compute that did not exist 18 months ago; and a cloud segment where H100 availability is better but still expensive enough to push serious workloads toward local ownership.

The GPU shortage did not end. It fractured into three distinct markets with different supply dynamics, different price points, and different bottlenecks.

For AI developers, the relevant question is not whether GPUs are scarce. It is which GPU segment matters for your workload, what it costs to own versus rent, and whether the new workstation tier changes the math on local inference. For a growing number of developers, the RTX PRO 5000 72GB has already changed that calculation. The card exists. It is available at workstation vendors. And it runs models that consumer hardware simply cannot.

The shortage narrative is a 2023 story. The real story is more interesting and more uneven. Some GPUs are easy to get. Others are not. The difference matters if you are the one trying to run something.