Generative AI cuts both ways for a studio. The tools are changing how creative work gets made, and the boom behind them is buying up the hardware the work runs on. The conversations everyone is having with Claude and ChatGPT run on cloud accelerators built from the same silicon and memory supply that render fleets draw on. Render hardware has no fabs or memory lines of its own. Anyone who tried to buy a graphics card in the mining boom or the pandemic shortage knows how this plays out: a buyer with more money and different priorities enters the market, and the part you budgeted for stops being available at that price, or at all.
The memory market is showing it first. Last October a 32GB DDR5 kit went for between $100 and $200. By this August, Tom's Hardware's price tracking had the cheapest kits starting around $350 and mainstream ones running roughly $400 to $600. Prices at least tripled in under a year, on one of the most ordinary parts in a computer.
Where the memory is going
Three manufacturers (SK Hynix, Samsung, Micron) control the vast majority of the world's DRAM, and per IEEE Spectrum they are shifting output toward high-bandwidth memory for AI accelerators, which sells for about three times the price of standard memory and can account for half the cost of a packaged GPU. Coming out of the last downturn, they added little new fab capacity through 2024 and most of 2025, so there is no slack to absorb the demand. Gartner's forecast has combined DRAM and SSD prices ending 2026 about 130% above 2025, enough in its estimate to cut worldwide PC shipments by 10.4%.
The pressure has already reached the cards themselves. PC Partner, the company behind Zotac and Inno3D graphics cards, said in its half-year results that entry-level GPU shortages will get "even more severe" in the second half of 2026. The graphics memory on those cards comes from the same DRAM industry that is prioritizing the AI parts.
Render farms are near the back of the line
Cloud rendering and AI run on largely the same hardware, and against the budgets of the tier-1 AI buyers, a render farm is one of the smaller customers in a line that has gotten long. The buyers ahead of you sign contracts no studio can match: IREN's SEC filings describe a roughly $3.4 billion agreement to deliver dedicated GPU capacity to NVIDIA over a five-year term. Once capacity is committed on that scale and that timeline, what breaks a delivery schedule is availability, and a higher hourly rate won't release a fleet somebody else contracted years ago. A bid that assumes this year's cloud GPU capacity at last year's spot price is carrying a risk line nobody wrote down.
I spend a good part of my week around the cloud vendors, and what I hear privately lines up with what you can now read in public.
The other shortage is power
A Guardian investigation by Aisha Down and Ed Zitron (published August 17, 2026) found that Microsoft had about 2.2 million AI chips installed, per internal documents, against roughly $280 billion of infrastructure spend since 2022. Microsoft disputes the arithmetic. Satya Nadella has described the risk in his own words: "you may actually have a bunch of chips sitting in inventory that I can't plug in" while the powered buildings to hold them get finished. Capacity stuck behind construction arrives whenever the buildings and substations do, which is a reason to expect the squeeze to ease unevenly rather than all at once.
The T4's shutdown date
In 2023 we rendered Refik Anadol's Coral installation for Davos: a render built from 100 million coral images, on up to 250 NVIDIA T4 GPUs across Google Cloud Spot VMs, delivered in three weeks against an estimated six on LA studio servers. The budget worked because T4 capacity was cheap and there was a lot of it. Google Cloud has since published the T4's end of support: as of August 1, 2027, "you cannot create, launch, or access any Google Cloud resources that run NVIDIA T4 GPUs," remaining instances get shut down, and no new three-year committed use discounts are offered. The P100 goes earlier, after September 15, 2026. The documented migration path is the G2 or G4 machine series, and the G2's L4 is a card Google's docs describe as purpose-built for AI inference. The compute that used to be cheap and everywhere for rendering now competes with the inference market for the same part.
What's a smart CTO to do?
Treat render capacity as a constrained input you book ahead, the way you already book crew or stage time. Four moves cover most of it:
- Book the capacity a delivery depends on. The clouds sell this literally now: Google Cloud's calendar mode reserves GPU capacity by start date and duration up to 90 days out, standard future reservations cover the render-class cards, and AWS Capacity Blocks take bookings up to eight weeks ahead. A reservation costs money whether you use it or not, and a committed use discount is not a reservation.
- Spread across zones and regions instead of assuming your usual one has depth. Spot pricing is public now: trackers like GPU Hour make it easy to compare regions and providers.
- Sort your workloads by which genuinely need current silicon and which only ran on a modern card because it used to be cheap. A denoiser and a distributed sim don't have the same claim on an L4.
- Feed fewer GPUs properly instead of starving more of them. Utilization is the lever most within your control, and we wrote up the feeding side in The invisible backbone.
The zone work takes real engineering. The rest takes planning time. The cost of skipping them shows up in delivery week, when you find out whether the depth you assumed in your usual zone was ever there.
On how long this lasts, the forecasts disagree. TrendForce sees memory contract prices still climbing through Q3 2026, and Intel's CEO has said there's no relief until 2028. Forecasts in this market have missed in both directions before, so for anything already on your calendar I'd assume prices stay high and capacity stays committed, and treat any easing as upside.
