Graphics cards, RAM, SSDs — what’s next?
The era of digital abundance, where any enthusiast could build a server at home to rival the capabilities of a small company, is coming to an end. Owning advanced hardware is increasingly becoming an elite privilege as chip prices soar and pre-order queues grow.
In a new piece, ForkLog explores why graphics cards have become a resource for the AI industry, how Nvidia has shifted its focus away from gamers, and why freelance designers are now renting power from cloud data centers. The key question we sought to answer is: how will the chip shortage impact blockchain decentralization, where SSDs and DRAM often play a crucial role.
Tech Feudalism or Temporary Hardships
Recently, based on the statements from leaders in the AI industry and memory chip manufacturers, it seems that the era of owning a powerful personal computer (PC) is gradually coming to an end.
There’s been active discussion in the media about a speech by Amazon founder Jeff Bezos from 2024, where he compared using a PC to having a generator during the era of centralized electricity supply. Some in the community view him as a prophet in this situation.
The latest hardware models are becoming the primary computational resource for training and maintaining large language models (LLMs). AI is depleting the inventories of HBM chip manufacturers, whose capacities were previously dedicated to consumer SSDs and RAM. As a result of rising component prices, the market could lose an entire class of budget devices this year.
In early February, TrendForce researchers raised their chip price forecasts. They expect a surge in contracts for consumer DRAM memory by 90–95% in Q1 2026 due to the AI segment boom, up from a previous forecast of 55–60%.
Moreover, training LLMs requires enormous amounts of data. The corporate sector has purchased stocks of SSDs with capacities starting from 2 TB and high write endurance. Silicon chip manufacturers, benefiting from the AI sector's higher revenues, are planning to reorganize their production capacities.
At the end of 2025, Micron Technology, a leader in memory chip production and previously a strong advocate for maintaining the desktop segment, announced the closure of its Crucial consumer line. Production will cease in Q2 2026 after nearly 30 years of the brand's existence.
Micron also plans to increase production of HBM chips, having invested $9.6 billion in new facilities in Hiroshima, Japan.
On February 12, Samsung Electronics announced the start of shipments of advanced HBM4 chips to unnamed clients. This move aims to close the gap with competitors in the critical components market for Nvidia's AI accelerators, including SK Hynix.
The world’s largest chip manufacturer finds itself in a challenging position: it is the main supplier of memory for Nvidia while also leading in the smartphone and consumer electronics segments. The company needs to maintain high-margin contracts in AI without compromising its position in gadget production.
In September last year, Samsung Semiconductor management attempted to balance the situation. The company confirmed that its GDDR7 memory production lines for high-end graphics cards can cater to both gamers and content creators as well as professional workstations.
These chips are used in Nvidia’s flagship gaming line, the GeForce RTX 5090. Launched in January 2025, this graphics card remains the undisputed leader, and the initially stated price of $1999 is now far from reality, with current offers ranging from $4000 to $5000.
The highly adaptive Chinese market is seizing opportunities as usual. According to Nikkei Asia, major Chinese memory manufacturers CXMT and YMTC are looking to significantly expand their capacities.
By 2027, they plan to launch factories in Shanghai and Wuhan, focusing primarily on DRAM and NAND rather than HBM, as market leaders do.
Alex Petrov, former CIO/CTO of Bitfury Group and co-founder of Hyperfusion, believes there’s no point in hoping for lower prices; it’s better to redistribute costs.
“There’s no point in waiting; we live in the here and now. If you need hardware for work, mining, or nodes, it’s better to buy now, accepting high prices, and allocate what you can temporarily do without. Deferred demand by 2028 could be enormous and unpredictable; we can only hope for old DDR3/4 and the release of new DDR6,” the expert shared in a comment to ForkLog.
Why Graphics Cards?
Why have graphics cards, which allowed players to enjoy Quake III Arena in 2000 and Fallout 4 in 2015, first been taken over by PoW mining and later absorbed by the AI industry? The answer lies in the specific capabilities of graphics accelerators, which can be better understood by comparing them to central processing units (CPUs).
A CPU is a genius capable of solving any type of software task: writing poetry, calculating taxes, managing an operating system. However, tasks are executed sequentially on each core.
In contrast, a GPU is like a factory with thousands of simple workers. Each is less intelligent than the genius but can operate simultaneously.
To render a frame in a game, the color of millions of pixels must be calculated. This equates to the same number of identical mathematical operations per second. The graphics chip was born for parallel computations.
A similar situation occurs in PoW mining with graphics cards. Mining is a kind of lottery where the device must generate random numbers billions of times per second to find the correct hash. GPUs were perfectly suited for this, leading to the first wave of their shortage before Ethereum transitioned to PoS in 2022.
Graphics processors have become a true boon for the AI industry. Modern LLMs like ChatGPT or Gemini are essentially gigantic tables of numbers (matrices). Their training involves endless multiplication of these matrices to determine “weights” (connections between neurons).
It turns out that the mathematics creating reflections on water in Cyberpunk 2077 is the same linear algebra that underpins neural network training. However, AI requires not only powerful computations but also colossal data transfer speeds. Ordinary gaming memory is insufficient for this — it has been replaced by expensive and scarce HBM, which tech giants are currently vying for.
Nvidia recognized this trend in time and, starting with the Volta architecture, began adding “tensor cores” to its graphics cards. These cores can simultaneously multiply matrices tailored specifically for AI tasks.
GPU on Demand and the Loss of Offline Capability
In the current situation, for at least the next two years, content creators, video editors, designers, gamers, programmers, AI architects, and anyone whose work critically depends on powerful hardware will have to make a choice: prioritize renting online resources or significantly overpay for a PC upgrade.
Given the shortage and backlog for certain components, demand for subscriptions is gaining momentum, forcing cloud data centers to become more client-oriented. There are companies offering flexible access to computing power and GPU rentals, such as Lambda Labs, Vast.ai, Hyperfusion, LeaderGPU, Hostkey, and others.
The RunPod service offers access to the scarce flagship RTX 5090 for $0.89/hour.
The Shadow platform provides a remote desktop without restrictions on running games and professional software for engineers and designers. Similar services like GeForce Now or Xbox Cloud do not offer such freedom but differ in pricing.
Even now, with a stable internet connection, a home smart TV can transform into a powerful workstation, provided the necessary hardware is ordered. This opens up previously inaccessible opportunities for many, but all responsibility for quality and uninterrupted operation shifts from users to the owners of tech parks, who may prioritize more important clients and comply with sanctions.
Petrov noted that data centers guarantee 24/7 availability, backup power, redundant connections, and proper maintenance quality.
“At the same time, you can store some things at home or work. It’s just often more expensive and less convenient,” he added.
According to him, many designers, video editors, producers, and artists are already being displaced by artificial intelligence. At a certain level, they have to turn to specialized AI applications that cannot be handled by “home hardware.”
“As the demands of LLMs grow exponentially, you can only keep small models on your phone or at home. Expert versions of larger sizes require a different scale, power, and speeds, which cloud data centers provide,” Petrov explained.
Bitcoin Leads Again
The entire IT sector relies on components, but for the blockchain industry, the chip shortage poses a real threat to decentralization and power redistribution.
“The rise in memory prices is a consequence of decisions made by individual commercial companies. Blockchain nodes are not the only ones affected; the prices of all devices with new DDR5 memory are rising: smartphones, PCs, everything. This also forces blockchains to become smarter and more economical, seeking various paths and solutions,” said the co-founder of Hyperfusion.
He pointed out the paradox of the current situation, where PoS networks are struggling:
“Proof-of-Stake has reduced mining energy consumption but shifted the burden from electricity to memory and disks for businesses and users. In a situation where components have increased in price by 3–5 times, PoS chains are caught in the ‘perfect storm’ of reality.”
In blockchains like Ethereum and Solana, the principle is “easy to create, but extremely expensive to verify.” Given that there are many nodes in the network and proofs take seven to nine steps, the barrier to entry for validators in PoS networks is often lower for deployment but higher in operational costs.
According to Petrov, in Ethereum, each node must keep the entire database of accounts, contracts, and balances readily accessible. This amounts to tens of millions of objects that are constantly updated. For fast operation, high-speed RAM and NVMe SSD combined in a RAID array are necessary.
Nodes must process each block. In high-frequency networks (Solana — 400 ms, Ethereum — 12 s), enormous resources are required for signature verification and transaction execution. In such systems, the requirements for full archival nodes are significantly higher: in Ethereum, an archival node requests 128GB RAM/at least 12TB SSD.
The decline in validator profitability due to rising component costs creates a new risk of blockchain centralization. In January, the daily number of active nodes in the Solana blockchain dropped to 800 — the lowest level since 2021. As support diminishes, it becomes increasingly difficult for small node owners to cover voting and infrastructure costs if they lack sufficient delegated stakes.
At the time of writing, the Nakamoto coefficient of the network has dropped to 19 (it was 33 in 2023).
The Ethereum Foundation is already discussing initiatives to lower the infrastructure entry threshold. In May 2025, Vitalik Buterin proposed EIP-4444, an update that could significantly reduce disk space requirements. It is expected that nodes will only store transaction history for the last 36 days while maintaining the current state of the network and the structure of Merkle trees. This approach allows for reduced storage volume without compromising the verification of the current state of the blockchain.
In the new realities of the “silicon curtain,” Bitcoin remains the “people’s blockchain.”
“In Bitcoin, there is no state verification, only UTXO, which can be easily cached. The PoW mining creation phase requires ASIC farms with enormous energy-efficient capacities, but validation remains extremely lightweight. Verifying the result in PoW is very simple and fast; that’s its beauty. The steps for a validating node are: get block data, check its hash, one or two hashing operations, compare target/difficulty, and it’s clear — yes/no,” Petrov explained.
For these reasons, a full Bitcoin node can operate even on a lightweight server or desktop, and sometimes even on new Raspberry Pi devices with 4–8 GB of RAM. The impact of memory shortages on PoW nodes is minimal. SSD prices are rising, but capacities up to 1 TB are still available, the specialist added.
What’s Next?
Petrov believes the era of personal hardware is not over. There are just different approaches and solutions for specific tasks:
“I love the quote ‘Cloud is someone else’s computer.’”
The industry is hastily seeking ways out of the chip crisis, developing new technologies:
- Magnetoresistive RAM (MRAM) is sufficiently non-volatile. It is about 1000 times faster than SSDs and more reliable than conventional memory. By 2026, it is expected to replace memory in critical systems (automotive, aerospace);
- CXL 3.1 (Compute Express Link) allows servers to “share” their RAM over the network. This is a lifesaver for data centers but ties users more closely to the cloud.
This current crisis is not the first in history but is the most structural. Previously, memory chips have faced similar challenges:
- 1986: The U.S. imposed an agreement on Japan, establishing a “price floor” for memory chips, leading to a tripling of DRAM prices in one year. American PC manufacturers (Commodore, Apple) nearly went bankrupt, and Intel left the memory market to focus on processors;
- 2011: Flooding in Thailand submerged Western Digital factories, producing 40% of the world’s HDDs. Prices skyrocketed by 190% and did not return to normal for two years.
The exponential growth of AI makes it difficult to realistically assess future market behavior. The anticipated launch of new capacities could alleviate the crisis by 2028, provided the current pace of development continues.
If AI agents become the foundation of the economy, demand for chips will grow faster than their production. In such a scenario, owning powerful PCs will become as elite a hobby as owning a racehorse. Whatever the future holds, remember to change your thermal paste on time.