Last month, in a glass-walled meeting room overlooking Vancouver's Coal Harbour, the founder of a decentralized compute network slid a hardware roadmap across the table and asked me the question that has quietly become the most important question in this sector: should we hedge our HBM supply?
Two years ago, nobody building a decentralized compute marketplace talked about high-bandwidth memory. They talked about GPUs. They talked about token incentives, about democratizing the data center, about routing around the hyperscalers. HBM was a supply-chain footnote belonging to TSMC and Samsung and SK Hynix — a line item in somebody else's earnings call.
Then Crypto Briefing — a crypto outlet, not a semiconductor trade publication — ran a short item reporting that Chinese AI chipmakers had raised prices by roughly 50% because of an HBM shortage. No company named. No baseline price. No quantified shortage data. No attributed source. Just a number and a cause, and the number was 50%.
That a crypto desk broke this story is not a curiosity. It is a signal. The AI compute wars and the crypto compute wars are converging on the same physical chokepoint, and the chokepoint is not the one the marketing decks describe. Code is law, but people are the soul — and apparently, so is physics.
Context: what HBM is, and why a crypto desk is writing about it
High-bandwidth memory is stacked DRAM. Multiple DRAM dies are layered vertically, connected by through-silicon vias, then bonded onto a logic die through 2.5D advanced packaging. The point of all that layering is bandwidth: an AI accelerator's compute units are useless if they are starved for data, and HBM is what feeds them fast enough to stay busy.
That physical reality has a financial consequence. In the bill of materials for a modern AI accelerator, HBM typically accounts for somewhere between 30% and 50% of total cost. This is the mechanical reason a memory shortage becomes a chip price hike: move the memory price and the whole card moves with it. There is no negotiating around arithmetic.
The supply side is brutally concentrated. SK Hynix holds roughly half the global HBM market, Samsung holds most of the rest, and Micron fills in a sliver at the edge. Three companies. That is the entire commercial universe for the component that determines whether an AI accelerator can be built at all. Ask yourself which part of that structure is decentralized.
Now add geopolitics. In December 2024, the US Bureau of Industry and Security extended export controls to cover HBM2E and above for Chinese buyers. This was the second chokehold. The first, applied years earlier, was advanced logic — the ability to fabricate cutting-edge accelerators at all. The second was memory. You can design a brilliant chip and still be unable to feed it.
Which is precisely why the story surfaced in a crypto publication. Mining farms have been pivoting to AI hosting for three years. DePIN compute networks — Akash, Render, io.net, Bittensor and their many imitators — are all selling the same scarce commodity the hyperscalers are buying. Tokenized compute, sovereign inference, decentralized training: every one of those narratives terminates, physically, in a memory stack that three firms build and one export-control regime can gate. The convergence is not thematic. It is literal. The same DRAM dies are being fought over by Nvidia, Huawei, and a token network in your portfolio.
Core analysis: the bottleneck nobody puts on the slide
The headline says HBM shortage. The real bottleneck is one layer deeper, and it is the layer that almost never appears in a pitch deck: advanced packaging.
HBM does not simply sit next to a logic die. It has to be stacked — DRAM dies connected by thousands of microscopic TSVs, micro-bumped or hybrid-bonded together, then integrated with the accelerator through a 2.5D interposer process of the CoWoS class. Every one of those steps is a manufacturing discipline with its own yield curve. You can have HBM dies and still fail, because you cannot package them.
This is where China's position is most exposed, and it is the detail the Crypto Briefing item never touched. The mainland has players in 2.5D packaging — JCET, Tongfu Microelectronics, Huatian — but their high-end capacity and yields do not approach TSMC's. So when the report attributes a 50% price increase to an HBM shortage, the honest technical reading is that the shortage is at least as much a packaging shortage wearing a memory costume. HBM is a four-part system: DRAM die, TSV stacking, advanced packaging, and substrate. China is behind on every part, and behind furthest on the parts that are least glamorous.
There is a second distinction the coverage blurred, and it matters more than the first. A shortage can mean two completely different things. It can mean the product is available but expensive — a market condition. Or it can mean the product cannot be bought at all — a strategic condition. The December 2024 controls put HBM2E and above into the second category for Chinese buyers. Once you see that, a 50% price hike stops looking like a market signal and starts looking like a euphemism. It is either rationing dressed as pricing, or a cost pass-through executed under duress. The first is inconvenient. The second is structural. The reporting never separated them, and that single omission changes the entire meaning of the event.
Scale the gap and the picture sharpens. Global leaders have been in mass production on HBM3E since 2024. China's domestic efforts — CXMT, Huawei's internal programs — are aiming for HBM2 and HBM2E-class output in meaningful volume somewhere in the 2026 to 2027 window, with HBM3 and beyond further out. That is a three-to-four-year lag on the single component that most determines whether an accelerator is competitive. Logic design is not the binding constraint. Memory bandwidth and packaging capacity are.
Here is where my own audit experience becomes uncomfortable. I have now sat in on four DePIN compute diligence calls this year. The pattern repeats with almost liturgical precision. The token incentives are elegantly designed. The emissions schedule is defensible. The hardware roadmap, without exception, is built on an implicit assumption that compute is fungible and supply is elastic — that if you need more, you add nodes, and if you add nodes, you get more.
That assumption is false at the physical layer. The bottleneck is not FLOPs. It is memory bandwidth and packaging capacity. A decentralized network cannot route around a CoWoS shortage by onboarding validators. It cannot mint its way past a TSV yield curve. Decentralization is a verb, not a noun — and the verb here is coordination, and coordination does not fabricate interposers. This is the part of the story that the crypto compute narrative has not internalized, and it is the part that should worry anyone underwriting these networks at current valuations.
There is a pricing-power paradox buried in the same number, and it cuts against the reflexive reading. When Chinese accelerators raise prices 50% and buyers accept, that is not evidence of weakness. That is evidence of pricing power — because the export controls removed the alternative. Nvidia's high-end parts are gated out of the Chinese market, so the domestic champion becomes the only option, and the only option sets the price. Geopolitics did not just constrain China's AI industry. It built it a walled garden with its own pricing regime. A scarcity premium is not a bug in that system. It is the moat.
The same logic applies uncomfortably well to crypto. When a tokenized compute network is the only sovereign option available to a buyer who cannot or will not touch centralized cloud, it can price above spot. That premium looks like strength. But it is manufactured by constraint, not by capability, and constraint can be engineered away by the same forces that created it. Investors who read scarcity premiums as durable advantages are reading the wrong variable.
The capital cycle adds a third turn. China's domestic buildout in HBM and advanced packaging is enormous — the third phase of the national semiconductor fund alone carries a registered capital in the hundreds of billions of renminbi, with memory and packaging explicitly among its targets. Capex like that does not stay capex. It becomes depreciation, and depreciation crushes gross margin. So the 50% price increase preserves margin today while the capacity buildout guarantees margin compression tomorrow. The hike and the buildout are working against each other on a two-to-three-year horizon. Anyone modeling these companies linearly is modeling a curve that bends against them.
Now translate that to the crypto compute thesis. GPU rental economics have two layers. The top layer is utilization — how well you schedule idle capacity. That is where decentralized networks genuinely compete, and they compete well. The bottom layer is hardware cost — and every GPU in the market got more expensive at the same moment, because the same HBM scarcity sits inside every accelerator regardless of who owns it. So DePIN's cost advantage compresses precisely when its narrative gets loudest. The mechanism that supposedly makes decentralized compute cheaper is downstream of a bottleneck it cannot influence.
This is why the bull market deserves suspicion here, not celebration. I have watched multiple compute networks raise nine figures this cycle on attestation claims that do not survive a supply-chain diagram. After my last audit, I started asking three questions in every diligence call, in this order. Where does the memory come from? Where does the packaging happen? What is the pass-through clause in the supplier contract? Most decks answer the first two vaguely and cannot answer the third at all. Trust is not verified on-chain. It is verified by a purchase order and an enforceable supply agreement, neither of which is trustless.
I learned this lesson the hard way, and not in semiconductors. Years ago I co-founded a community fund whose treasury was drained through a flawed multisig — not because the code was wrong, but because the governance around the code was. I spent the following years convinced the fix was better code. It never was. The fix was understanding the layers beneath the abstraction, the places where a system's elegance hides its fragility. HBM is that layer for the entire compute economy, and crypto has spent a decade treating it as somebody else's problem.
The contrarian angle: the coverage got the causality backwards
The consensus reading of the 50% hike is that China's AI industry is being hurt. The less comfortable reading is that the hike is the first public evidence that Chinese accelerators have attained captive-market pricing power at home, and that export controls have produced the opposite of their advertised effect. The controls did not stop Chinese AI. They guaranteed it would run on a slower, costlier, fragmented stack — while its domestic champions extracted rents inside that stack. That is not a defeat for the Chinese industry. It is a toll booth.
But the point that should genuinely unsettle a crypto reader is different. Every decentralized compute network I have reviewed in the last eighteen months has been selling a decentralization story whose physical layer is more centralized than its governance layer. Three foundries. Three memory makers. One or two advanced-packaging players. The voting mechanism may be on-chain and permissionless. The assembly line is neither. We spent a decade decentralizing consensus and forgot to decentralize the supply chain, and the bull market is perfectly happy to fund the narrative while skipping the audit. Decentralization of the ledger is not decentralization of the machine.
Takeaway: three signals worth tracking
Watch three things over the next year. First, whether CXMT or Huawei confirms volume HBM shipments — that tells you when China's memory wall starts to move. Second, whether the Bureau of Industry and Security widens the HBM controls again, because that would convert a shortage into a severance. Third, and most telling for this audience, whether any DePIN compute network publishes its memory supply chain and pass-through terms. The first two tell you where the physical capacity goes. The third tells you whether crypto's compute story is engineering or marketing. Decentralization is a verb. Verbs require materials, and the materials are stacked in three factories that belong to somebody else.
