Price with confidence
Compute becomes a fixed cost. Your margins don’t move when a provider reprices.
Our thesis · October 2026
Lock in your compute costs while you still can.
Compute is the biggest cost in an AI business. Memory stays short until 2028, the cheapest rentals need three-year contracts, and new GPUs are rationed. Owning fixes that cost for the life of the hardware.
- DRAM undersupply forecast for 2027 (Goldman Sachs)
- −5.9%
- NAND undersupply forecast for 2027 (Goldman Sachs)
- −4.6%
- earliest the memory squeeze eases, per Goldman
- 2028
- 3-year rental of one 8-GPU H200 node. You own nothing after.
- $471k
The squeeze
Memory is short, and AI is what’s using it up.
Every GPU runs on high-bandwidth memory, and every server on DRAM and flash. Goldman Sachs sees DRAM undersupply widening from 5.0% in 2026 to 5.9% in 2027, the worst since 2017, with NAND 4.6% short. Short supply means higher prices, for new servers and for every rented hour priced off them.
Hardware you already own was bought at yesterday’s price. Rented capacity is repriced at every renewal.
The contract trap
The cheap rental rate is a 3-year commitment.
Indian GPU clouds save their best prices for long terms. Neysa lists an 8-GPU H200 node at $16,758 a month, or $13,084 if you sign for 36 months. That’s about $471k over three years.
All the commitment of owning, none of the asset. If you’re signing for three years anyway, buy the servers. Same horizon, and at the end they’re yours to run, lease out or sell.
Your business model
Fix your cost per token, per render, per run.
Front-load the tax benefit
Depreciate 40% of the server in year one. Rent is deducted only as you pay it. The schedule
Keep the upside
Lease out idle time. Sell and trade up when the next GPU lands.
Supply
Capacity goes to those who already have it.
NVIDIA has closed new H100 orders. B200 runs 8–16 weeks with an OEM allocation; 30+ without. In a shortage, the teams that keep shipping are the ones whose GPUs are already racked.
Current lead timesThe rule
Own the baseline. Rent the peaks.
Buy the GPUs you run every day. Host them with us at Yotta. Rent only for launches, experiments and spikes. That’s the steadiest way to run an AI business through 2028.
Questions
Will GPU prices go up in 2027?
Goldman Sachs expects DRAM and NAND undersupply to deepen in 2027 and last into 2028, which supports higher memory prices. GPUs, servers and rented capacity all depend on that memory.
Is it cheaper to buy GPUs or sign a long-term rental?
If you are committing for three years anyway, compare the total contract against buying and hosting the same servers. A 36-month rental is a three-year commitment that leaves you with no asset at the end.
When is renting still the right call?
Short projects, experiments and traffic spikes. Rent the peaks; own the baseline you run every day.
Sources
- Benzinga: Goldman Sachs sees the memory squeeze running to 2028 (June 2026)
- Yahoo Finance: memory shortage deepens to worst levels since 2017
- TechFlow: Goldman Sachs memory report analysis (June 2026)
- Neysa: NVIDIA H200 price in India, rent or buy (Sept 2026)
- Compute Exchange: reserved vs on-demand GPUs in 2026
