The ledger does not lie, only the operators do. SK Group Chairman Chey Tae-won’s recent assertion that memory chip demand will surge 50-60%—with AI-driven HBM demand up 100%—is not a bullish thesis; it is a structural confession. The semiconductor supply chain, long optimized for desktop cycles, now faces a collision with the exponential appetite of large language models and, by extension, the crypto networks that depend on them. Silence in the code is a bug waiting to happen, and the silence from HBM suppliers is deafening.
Context: The Unspoken Link Between HBM and Web3
This analysis begins not at the chip fab, but at the blockchain consensus layer. Every AI training run, every zk-SNARK proof generation, and every on-chain inference request consumes memory bandwidth. The recent surge in AI-agent protocols—from autonomous trading bots to decentralized physical infrastructure networks (DePIN)—has created a parallel demand curve that mirrors centralized AI clouds. SK Hynix, the industry leader in High Bandwidth Memory (HBM), is the gatekeeper. Chey’s comments, reported by Maeil Business Newspaper, signal that even with $20 billion in committed capex, the company cannot build fast enough. For crypto projects reliant on GPU clusters (e.g., Render Network, Livepeer, or zk-rollup sequencers), the bottleneck is not just compute—it is memory.
This is not a future risk; it is a present constraint. In Q3 2024, SK Hynix’s HBM3E revenue grew 300% YoY, yet lead times for advanced HBM packages remain above 12 months. Chey explicitly stated that equipment, personnel, and construction cycles are the binding constraints—not technology or demand. This admission contradicts the market’s implicit assumption that supply can scale linearly with demand. For crypto mining and AI-Rollup infrastructure, this means that the cost of memory will remain structurally high, squeezing margins for any project that requires high-bandwidth memory for proof generation or large-scale model inference.
Core: A Systematic Teardown of Chey’s Supply-Demand Thesis
Let us dissect the numbers. Chey forecasts total memory demand growth of 50-60%. He then predicts AI-specific memory (HBM) growth of 60-100%. He further claims the gap between supply and demand could widen. To validate this, we must examine the physical capacity constraints.
First, wafer fabrication. SK Hynix’s upcoming M15X fab in Yongin will produce DRAM on the 1bnm node. The cleanroom build-out takes 24 months. The key equipment—ASML’s EUV scanners—has a 12- to 18-month delivery backlog. Chey’s assertion that “capacity cannot ramp fast enough” is mathematically sound. A 25% increase in HBM bit supply over the next year would require an 8% increase in DRAM wafer starts, plus an even larger increase in TSV (through-silicon via) packaging capacity. TSV equipment, primarily from Tokyo Electron and Applied Materials, has similar lead times. The result: total HBM bit output in 2025 will likely be 300-400 million GB, while demand (from NVIDIA alone) is estimated at 500+ million GB. That is a clear structural deficit.
Second, the packaging bottleneck. HBM requires advanced packaging—CoWoS or equivalent. SK Hynix has outsourced some TSV to ASE and Powertech, but capacities are limited. Chey noted that “even if every factory runs full, it will not be enough.” This aligns with third-party estimates that CoWoS capacity will only grow 40% in 2025, against 80% HBM demand growth. For crypto projects, this means that GPU cluster expansion will be capped not by GPU availability (which is also constrained), but by the speed at which memory modules can interface with them.
Third, the economic incentive. Chey urged competitors to “build, not restrict supply.” This is a strategic call to avoid the classic oligopoly trap of collusive output reduction. Instead, he advocates scaling the entire pie. From a game-theory perspective, this is rational only if the demand curve is sharply upward-sloping and inelastic. Historical data from 2018–2023 shows that memory cyclicality was driven by over-investment followed by crashes. Chey is betting that AI demand is structurally different—and the weight of evidence from NVIDIA’s order book supports him. But there is a hidden assumption: that AI demand growth is not a bubble. If a correction occurs, the excess HBM capacity will collapse prices, devastating the semiconductor industry.
For the blockchain sector, the implication is stark. Projects that depend on affordable high-bandwidth memory—such as zk-proof accelerators (e.g., Ingonyama’s prover boards) or decentralized AI compute marketplaces—will face cost inflation of 30-50% over the next two years. This will force a natural selection: only tokens with strong profitability or subsidies (e.g., Render’s burn-and-mint model) will survive. The rest will be priced out of the market.
Contrarian Angle: What the Bulls Got Right
Despite the grim supply outlook, there is a counterintuitive opportunity. Chey’s prediction of a widening gap implies that the price of memory will remain elevated—or even rise. This is not universally bearish for crypto. For protocols that build on top of this scarcity, such as memory-leasing markets (e.g., Akash’s planned memory-focused deployment), higher ASPs mean higher revenue per resource unit. In a bull market for AI-crypto, the demand for compute and memory can outstrip supply, raising utilization rates and rewarding early capacity holders.
Furthermore, Chey’s call for accelerated capex might trigger a positive feedback loop: as memory companies invest in new fabs, they will eventually bring online massive new capacity, which will later become surplus. Historically, memory oversupply leads to price crashes, which then enable cheaper deployment cycles. For crypto projects that can time their hardware procurement to the eventual trough (expected 2027–2028), the long-term cost basis may fall dramatically. The bulls who extrapolate cheap memory from five-year cycles may be correct, but only if they can survive the next two years of high prices.
Another point the bulls got right: the diversity of demand. Chey explicitly mentioned that “the customer base is widening beyond NVIDIA.” Chinese AI chip makers, hyperscalers (AWS, Google, Microsoft), and even automotive AI are all soaking up HBM. For blockchain-based GPU networks, this diversification reduces the risk of a single-point collapse. If NVIDIA stumbles, other buyers will absorb the memory, providing a pricing floor. The crypto networks that have contracts with multiple hyperscaler partners—like Render’s alliance with Google Cloud—will be partially hedged.
Takeaway: The Chain Will Remember Who Bought the Dip
Proof is cheaper than trust, yet still ignored. Chey’s data should be a wake-up call for crypto asset managers. In a world where HBM supply growth is structurally capped at 40% while demand grows at 70%, the cost of running any memory-intensive blockchain operation will rise faster than the price of native tokens. The protocols that will survive are those that pre-negotiate long-term memory contracts, integrate dynamic pricing (like Filecoin’s storage market), or focus on compute-light workloads (e.g., simple transaction verification) that run on low-power DDR5 instead of HBM.
The ledger of semiconductor capacity does not negotiate; it only confirms. If you are building a DePIN project that requires high-bandwidth memory, your business model must account for a 40%+ cost increase over the next 18 months. Failing to do so is not a strategy—it is a bug.
Consensus is not a feature; it is the foundation. The industry consensus, as voiced by Chey, is that the HBM shortage will persist. The price of that consensus will be paid by those who ignored the supply chain. History is the only reliable audit trail, and it shows that memory shortages always end in one of two ways: a price spike that kills marginal demand, or a capacity glut that crashes margins. The current trajectory points to the former. Crypto projects should prepare accordingly.