The Graphics Card Is Disappearing — and AI Is Why

I made a little pocket money flipping graphics cards this year.

In March, I bought a 5070 Ti laptop for 14,200 yuan, used it for a couple of months, and sold it for 17,000. In July, I bought a 5070 Ti desktop for 13,800, used it for about a month, and sold it for 15,500. In October, after seeing news that HBM prices would rise again, I bought a TUF 5070 Ti for 8,200 and sold it three days later for 8,500. It was in good condition, so I didn’t push the price.

The three trades made me less than 5,000 yuan. What I earned was a few crumbs that fell off when AI compute pushed consumer graphics cards to the back of the queue.

Start with one number

In Nvidia’s latest quarterly report, data center revenue accounted for about nine-tenths of total revenue. Gaming was a single-digit percentage.

That number would have been unimaginable five years ago. In 2020, gaming was still roughly half of Nvidia’s business. Today, it is barely a rounding error.

The ratio matters less than the logic of capacity allocation behind it.

TSMC’s wafer capacity is finite. CoWoS advanced packaging capacity is finite. HBM supply is finite. Every additional B200 Nvidia produces consumes packaging resources, memory resources, and board-level capacity. Those resources could otherwise have gone into producing 5090s.

Jensen Huang is not a fool. A single B200 sells for tens of thousands of dollars. A rack-level solution can sell for hundreds of thousands, even over a million yuan. The customers are Microsoft, Amazon, OpenAI, Oracle. A 5090 sells for thirty, forty, maybe fifty thousand yuan. The customers are gamers and a small number of enterprise AI deployment teams. If you were Jensen Huang, who would you supply first?

Nvidia is not asking whether it wants to take care of gamers. It is allocating scarce resources. The answer is in the financial statements. You don’t have to guess.

The HBM chain reaction

Consumer graphics cards do not use HBM. The 5090 uses GDDR7. Many people assume HBM price increases have nothing to do with gaming cards.

Wrong.

HBM, server DDR5, and GDDR7 compete for the same things: wafers, advanced packaging, equipment, and capital expenditure. Samsung, SK Hynix, and Micron direct their best capacity toward HBM and server DRAM because that is where the margins are highest. The supply growth of consumer DDR5 and GDDR7 is therefore suppressed.

This is not just about switching a production line. It is about capital expenditure priorities. Building a new fab takes years. Buying equipment requires queuing. Ramp-up and yield improvement take time. When memory makers expand capacity, they expand HBM first. Consumer-grade memory comes later in the queue.

So even though the 5090 does not use HBM, its memory costs are still pushed up by this round of AI-driven memory demand. When BOM costs rise, graphics card prices cannot stay unaffected.

Enterprise AI deployment is also taking cards

Another variable is easy to overlook: many Chinese companies building AI agents will not buy B200 or B300 servers. The initial investment of several million yuan is too much for many of them.

They buy 4090Ds. Some even buy 5090s through informal channels.

Enterprise procurement looks at compute and VRAM, not whether “this card was designed for gaming.” When 4090Ds and 5090s are bought in batches by enterprises, the supply available to gamers shrinks. This variable did not exist in previous graphics card cycles, or at least was not mainstream. Now it is normal.

Capacity expansion is not a switch

Some will say Samsung, Micron, and SK Hynix are all expanding capacity. Once capacity comes online, prices will naturally fall. The memory industry has always been cyclical.

That is half right.

Expansion is happening. But from fab construction to production, from equipment procurement to yield ramp-up, it takes years. And when new capacity is released, HBM and server DRAM get served first. Consumer GDDR gets allocated later in the queue.

Demand matters too. Suppose memory capacity increases by 30% over the next two years, but memory demand from global AI compute construction grows even faster. Even with capacity expansion, supply and demand can remain tight. If AI capital expenditure suddenly slows, the situation becomes completely different.

There is another question: after the previous memory price collapse, will Samsung, Micron, and SK Hynix really expand at any cost and drag themselves back into a loss-making cycle? They have memory. More likely, they will expand in a disciplined, selective way, prioritizing high-margin products rather than indiscriminately starting a price war.

Memory prices will not rise forever. The cycle will eventually turn. But when it turns does not depend on manufacturers announcing expansion. It depends on whether newly released capacity can exceed growing demand.

Graphics cards are becoming a “secondary business”

In the past, we assumed graphics card price increases were temporary: mining booms, shortages, channel hoarding. Give it a while and things would return to normal.

Now we need to accept a different reality: consumer graphics cards are shifting from the core business of the GPU industry to a secondary business under the rapid development of AI.

“Secondary” does not mean “abandoned.” GeForce still has ecosystem value for Nvidia: developers need it, game engines need it, CUDA adoption needs it, brand and cash flow need it. Nvidia will not kill gaming cards. But it will prioritize data centers when allocating wafers, packaging, memory, and board capacity. AMD is the same.

It is like a company that will not shut down an old product line that still makes a profit, but will put its best engineers, best capacity, and most capital expenditure into the fastest-growing new business. Consumer graphics cards are in that position today.

Will prices keep rising in 2027?

My judgment: there is a high probability that graphics card prices will continue to rise in 2027, but the exact magnitude is hard to say. A more careful formulation: from late 2027 to 2028, gaming graphics cards may stabilize at a high level. Prices will only begin to fall when the 60-series launches and the volume of cards on the market clearly increases.

But this judgment is conditional. If AI capital expenditure continues to beat expectations, HBM and DRAM capacity release lags demand, and gaming GPU production continues to be squeezed, then consumer graphics cards will likely remain at high levels in 2027.

There are also many countervailing variables: AI capital expenditure suddenly cools; memory makers expand faster than expected and the DRAM cycle turns early; export controls change, altering the supply of China-specific cards and grey-market cards; TSMC releases more advanced process and packaging capacity; gaming demand weakens and second-hand supply increases; if 60-series pricing is very high, older cards will also fall slowly.

So “prices will keep rising in 2027” is a probability judgment, not a certainty. The memory cycle never rises forever. The turning point depends on the race between AI demand growth and capacity release.

The money I made

Back to those three trades.

Buy at 14,200, sell at 17,000. Buy at 13,800, sell at 15,500. Buy at 8,200, sell at 8,500. Each was a low-buy, high-sell trade. Each landed on the news cycle of HBM price increases.

I was not being smart. A cycle handed out an information-gap red envelope.

In the short term, it looks like arbitrage. In the long term, it is a trade on the trend of “AI compute squeezing the consumer supply chain.” As long as the trend holds, the red envelope exists. When the trend turns, it disappears. Second-hand cards also carry warranty risk, condition risk, fake or refurbished card risk. Liquidity can suddenly dry up at high prices. At larger volume, there are tax and compliance issues.

Treat it as a cyclical trade, not risk-free arbitrage.

Conclusion

Discussing graphics card prices today can no longer be done only from the perspective of gamers.

The AI compute boom has pushed up HBM, changed global DRAM capacity allocation, and squeezed the supply of consumer DDR5 and GDDR. Nvidia and AMD are directing their best resources toward data centers. Enterprise AI deployment is taking 4090Ds and 5090s off the secondary market. Consumer graphics cards are gradually shifting from the core business of the GPU industry to a secondary business under the rapid development of AI.

This trend will likely continue. But the specific price path depends on four variables: the pace of AI capital expenditure, the rhythm of memory capacity release, export controls, and 60-series pricing.

Graphics cards will not rise forever. The memory cycle will not rise forever either. The main line of this cycle is no longer gamers. It is AI infrastructure.

And those of us flipping a few cards in the secondary market are only earning a few crumbs from this big cycle. How long the crumbs keep falling does not depend on us. It depends on when the machine above changes gears.