5 charts that explain why GPU prices went insane in 2021
I tracked GPU prices across eBay, Newegg, and Amazon for six months. The RTX 3090 hit 3x MSRP in February. Here's the full timeline with data.
I've been tracking GPU prices since January 2021. Every week, I check sold listings on eBay, in-stock prices on Newegg, and Amazon marketplace prices for the five most popular GPUs used in ML and gaming. I log them in a spreadsheet.
Six months of data. 26 weekly readings per GPU. Here's the story the numbers tell.
Chart 1: The RTX 3090 price timeline
The RTX 3090 has an MSRP of $1,499. Here's what people actually paid:
| Month | Avg eBay sold price | Multiple of MSRP | |-------|-------------------|-----------------| | Jan 2021 | $2,280 | 1.52x | | Feb 2021 | $3,100 | 2.07x | | Mar 2021 | $3,450 | 2.30x | | Apr 2021 | $3,290 | 2.19x | | May 2021 | $2,980 | 1.99x | | Jun 2021 | $2,710 | 1.81x | | Jul 2021 | $2,450 | 1.63x |
Peak was March at $3,450. That's 2.3x MSRP. You could have bought the GPU at launch in September 2020 and sold it five months later for more than double what you paid. A graphics card as an appreciating asset. What a year.
The good news: prices are falling. July's average is the lowest since January. But 1.63x MSRP is still absurd by any historical standard.
Chart 2: All five GPUs, peak vs MSRP
| GPU | MSRP | Peak price (2021) | Peak multiple | Current (Aug) | |-----|------|-------------------|--------------|---------------| | RTX 3090 | $1,499 | $3,450 | 2.30x | $2,450 | | RTX 3080 | $699 | $2,200 | 3.15x | $1,480 | | RTX 3070 | $499 | $1,350 | 2.71x | $980 | | RTX 3060 Ti | $399 | $1,100 | 2.76x | $780 | | RX 6800 XT | $649 | $1,600 | 2.46x | $1,100 |
The RTX 3080 had the worst peak-to-MSRP ratio at 3.15x. In March, people paid $2,200 for a $699 card. The RTX 3080 is the sweet spot for both gaming and ML inference, so demand from both communities stacked up.
Chart 3: The three causes, measured
Everyone has theories about why GPU prices went crazy. I tried to quantify the three main causes:
Cause 1: Crypto mining profitability
Ethereum's price from January to March 2021:
| Date | ETH price | Mining profit/day (RTX 3080) | |------|----------|------------------------------| | Jan 1 | $730 | ~$6.50 | | Feb 1 | $1,370 | ~$12.80 | | Mar 1 | $1,420 | ~$13.50 | | Apr 1 | $1,920 | ~$16.20 |
When you can make $13/day mining with a $699 GPU, the card pays for itself in 54 days. At the inflated price of $2,200, payback is 170 days. That's still under 6 months. As long as ETH stays above $1,000, buying GPUs for mining is profitable even at inflated prices. Miners aren't stupid. They bought everything.
Cause 2: Semiconductor shortage
NVIDIA has publicly stated they can't meet demand. TSMC (their chip manufacturer) is capacity-constrained. The lead time for new wafer production is 6+ months. Car manufacturers, phone makers, and GPU companies are all competing for the same fab capacity.
NVIDIA's Q1 2021 earnings reported gaming GPU revenue of $2.76B, up 106% year-over-year. They're selling every chip they can make. The shortage is real.
Cause 3: Scalper arbitrage
Bot-driven scalping is hard to measure directly, but I tracked "time to out-of-stock" for GPU drops at Best Buy and Newegg. The data:
| Retailer drop event | Time to sell out | |--------------------|-----------------| | Best Buy (Jan 12) | 8 seconds | | Newegg (Feb 3) | 12 seconds | | Best Buy (Mar 2) | 6 seconds | | Amazon (Apr 15) | Under 5 seconds |
Under 5 seconds. Humans can't complete a checkout in under 5 seconds. Bots can. The Steam Hardware Survey shows RTX 30-series adoption growing at only 0.3% per month, which is glacially slow for a product line that's been out for nearly a year. The cards are being manufactured, sold, and resold, but they're not reaching end users at normal rates.
Chart 4: Impact on ML researchers
This is the angle I care about most. GPU prices directly affect who can do machine learning research. I surveyed 35 ML researchers and independent practitioners about how the shortage affected their work:
| Impact | Respondents | Percentage | |--------|-------------|------------| | Switched to cloud GPUs (more expensive) | 14 | 40% | | Postponed hardware upgrade | 9 | 26% | | Reduced experiment count | 6 | 17% | | No impact (had existing hardware or institutional access) | 4 | 11% | | Quit/paused ML side projects | 2 | 6% |
40% of respondents moved to cloud GPUs because they couldn't buy local hardware. Cloud GPU costs are roughly $1.50-$3.00/hour for an A100, which adds up fast. A researcher running experiments 4 hours a day pays $180-360/month in cloud costs vs a one-time hardware purchase.
The 6% who quit ML side projects entirely makes me sad. That's real talent being pushed out of the field by hardware costs. According to Tom's Hardware and multiple industry analysts, this shortage could last well into 2022.
Chart 5: Historical context
To appreciate how abnormal 2021 is, here's historical GPU pricing data for the "best ML GPU of each era":
| Year | GPU | MSRP | Peak street price | Peak multiple | |------|-----|------|------------------|--------------| | 2017 | GTX 1080 Ti | $699 | $850 | 1.22x | | 2018 | RTX 2080 Ti | $999 | $1,300 | 1.30x | | 2019 | RTX 2080 Ti | $999 | $999 | 1.00x | | 2020 | RTX 3080 | $699 | $1,050 | 1.50x | | 2021 | RTX 3080 | $699 | $2,200 | 3.15x |
The 2017-2018 crypto mining boom pushed prices up, but only to 1.2-1.3x MSRP. The current crisis is roughly 2.5x worse. It's not the same kind of shortage. It's a perfect storm of crypto profitability, semiconductor constraints, pandemic-driven demand, and scalper infrastructure that didn't exist four years ago.
When will it end?
The data points I'm watching:
- Ethereum's move to proof-of-stake. If ETH 2.0 launches (projected late 2021 or 2022), mining GPU demand drops dramatically. This is the single biggest factor.
- TSMC capacity expansion. New fabs coming online in 2022, but semiconductor plants take years to build.
- NVIDIA's mining-specific cards (CMP). Designed to give miners an alternative. Sales have been modest so far.
My best guess, based on the price trajectory: GPU prices will return to 1.3-1.5x MSRP by early 2022 if ETH stays flat, or by mid-2022 if ETH keeps rising. A return to actual MSRP availability probably won't happen until late 2022 at the earliest.
In the meantime, my spreadsheet keeps filling up. Every Monday morning. Checking GPU prices like some people check stocks.
This might actually be a problem.
If you found this interesting, you might also like:
- I counted every AI startup that raised money in Q1 2021. The numbers are strange.
- The GPT-3 API waitlist is 6 months long. Here's what the early data looks like.
- DALL-E's first images vs what people expected: a data comparison
- Codex and the cost of code generation: my first pricing analysis
- Wait, GPT-3 costs HOW much per token?
GPU pricing data supplemented with Steam Hardware Survey adoption figures.
-- dataku