On June 12, 2024, JPMorgan’s chief semiconductor strategist released a note that, on its surface, was about stock valuations. But beneath the PE multiples and earnings projections lay a starker truth: the AI chip supply chain is structurally broken, and it won’t be fixed for years. The strategist argued that “meaningful supply growth won’t materialize until 2028,” a statement that sent ripples through the investment community. For blockchain, this is not a mere stock market signal—it is a direct threat to the availability of the very hardware that powers proof-of-work mining, decentralized AI inference, and zero-knowledge proof computation.
I have spent 27 years dissecting financial systems and blockchain protocols, from auditing MakerDAO’s collateral thresholds to modeling Terra’s death spiral. When I see a bottleneck that cascades across industries, I treat it as a systemic risk. This article is not a commentary on NVIDIA’s share price. It is a forensic teardown of the semiconductor supply chain, viewed through the lens of blockchain’s hardware dependency. We will examine the technology, the geopolitics, and the hidden assumptions behind the “2028” deadline, and answer the question every blockchain builder should be asking: What happens if the chips don’t come?
Context: The JPMorgan Thesis and Its Blockchain Implications
JPMorgan’s strategist, whose note I analyzed using my seven-dimensional framework, argued that AI semiconductor stocks would rebound in the summer of 2024, driven by sustained earnings growth. The core reasoning: demand for AI training chips (GPUs) and high-bandwidth memory (HBM) remains insatiable, while supply expansion is constrained by bottlenecks in advanced packaging (CoWoS) and extreme ultraviolet (EUV) lithography equipment. The strategist explicitly stated that “substantial supply growth will not occur until 2028,” meaning the current shortage is not a cyclical blip but a multi-year structural deficit.
For blockchain, this is critical. The most visible overlap is proof-of-work mining, where ASIC chips (e.g., Bitmain’s Antminer S21) compete with AI chips for the same leading-edge fabrication nodes (5nm/3nm). But the impact goes deeper. Decentralized AI networks like Render Network, Akash, and Gensyn rely on idle GPUs for rendering and inference. If GPU supply remains tight, these networks either face higher costs or slower adoption. Furthermore, zero-knowledge proof generation (used by Layer 2 rollups like zkSync and StarkNet) is computationally intensive and benefits from high-performance chips. A prolonged supply crunch means higher operational costs for validators and sequencers.
The strategist’s confidence was high—he assigned a 9/10 to demand analysis—but his assessment of technology details was weak (3/10). As a due diligence analyst who has audited both code and supply chains, I find this asymmetry dangerous. The bullish case rests on three fragile assumptions: (1) TSMC can scale CoWoS packaging fast enough, (2) hyperscaler capital expenditure (capex) will not slow, and (3) export controls will protect incumbent margins rather than triggering retaliatory disruptions. Let’s examine each through the lens of blockchain’s hardware dependency.
Core: The Systemic Fragility of the AI Chip Supply Chain
1. The Packaging Bottleneck: Why CoWoS Matters for Blockchain
JPMorgan’s core insight—that supply growth is constrained until 2028—hinges on a single technology: chip-on-wafer-on-substrate (CoWoS), TSMC’s advanced packaging solution. AI GPUs like NVIDIA’s H100 and B100 use CoWoS to stack high-bandwidth memory (HBM) alongside the logic die. Without CoWoS, the chip cannot function. The problem is that CoWoS capacity is extremely limited. TSMC is spending billions to double capacity by 2025-2026, but even then, demand—driven by AI companies and hyperscalers—will outstrip supply. The strategist’s 2028 timeline reflects the reality that a new packaging plant takes 2-3 years to build and another year to ramp to full yield.
How does this affect blockchain? Consider Bitcoin mining ASICs. These chips are designed on 5nm/7nm nodes and packaged using traditional methods, not CoWoS. However, they compete for the same front-end capacity at TSMC. If AI chips consume the bulk of 5nm and 3nm wafers, mining ASIC manufacturers face longer lead times and higher prices. Bitmain’s Antminer S21, launched in late 2023, already saw delayed deliveries. A sustained AI supply crunch could push new mining hardware availability to 2026 or later, compressing miner margins and potentially centralizing hashrate among those with early access to fabs.
But the deeper blockchain impact is on decentralized AI networks. Projects like Render and Akash aggregate consumer-grade GPUs (e.g., RTX 4090) for rendering and inference. Consumer GPUs are fabricated on older nodes (8nm/12nm) and are less affected by the CoWoS bottleneck. However, if AI companies pivot to using consumer GPUs as a stopgap, they will bid up prices. In 2023, NVIDIA’s RTX 4090 saw price spikes of 30-50% due to AI demand. Decentralized GPU networks rely on a stable supply of affordable cards. A protracted shortage would reduce the economic incentive for individual miners to join these networks, slowing their growth.
2. The EUV Equipment Trap: A 18-Month Lead Time That Cannot Be Shortened
The strategist noted that “key equipment delivery cycles remain extended.” Specifically, ASML’s EUV lithography machines—essential for 5nm and below—have a lead time of 12-18 months. High-NA EUV machines, needed for 2nm, are even longer. This is not a supply chain glitch; it is a physical limitation. ASML builds only a handful of these machines per year, and they are pre-allocated to TSMC, Samsung, and Intel. The lead time cannot be compressed without redesigning the machines.
For blockchain, this means that any new fab capacity—whether for AI chips or mining ASICs—cannot come online faster than the EUV delivery schedule. The strategist’s 2028 estimate is consistent with the fact that a new fab ordered today would take 3-4 years to receive its first EUV tool and another year to ramp. Consequently, any blockchain project that depends on leading-edge chips (e.g., specialized zero-knowledge proof accelerators) must assume that hardware will remain scarce and expensive for at least four more years.
I recall auditing the Zilliqa whitepaper in 2017. They claimed sharding would enable linear scalability, but I found a critical edge-case in transaction finality. Similarly, many blockchain projects today assume hardware will become cheaper and more abundant over time. The JPMorgan analysis suggests the opposite: hardware costs will remain elevated, and lead times will stretch. Complexity hides risk. Projects building on the assumption of infinite GPU supply are, in my view, ignoring a systemic fragility.
3. Hyperscaler Capex: The Pendulum That Swings Markets
The strategist’s demand thesis relies on hyperscalers (Microsoft, Amazon, Google, Meta) maintaining or increasing capital expenditure. These four companies are spending over $200 billion annually on data centers, a significant portion on AI chips. JPMorgan assumes this spending will continue, citing the “AI arms race.”
From my experience analyzing MakerDAO’s collateralization ratios in 2020, I learned that assumptions about liquidity—financial or computational—can reverse quickly. A slowdown in hyperscaler capex would trigger a cascade: demand for AI chips would drop, TSMC would reallocate capacity to other customers (including mining ASICs), and the “shortage” would evaporate. But such a slowdown would also likely coincide with a recession, reducing demand for blockchain services overall. The risk is that the JPMorgan thesis is a double-edged sword: as long as capex stays high, chips are scarce; if capex falls, chips become available but the economic environment deteriorates.
4. Geopolitical Protectionism: The Hidden Premium
The strategist did not explicitly discuss geopolitics, but my analysis scored it an 8/10 in importance. US export controls on advanced chips to China have created a bifurcated market. Chinese AI companies cannot purchase NVIDIA H100s, so they buy alternative chips (e.g., Huawei Ascend) or smuggle them. This reduces total available supply for the rest of the world, inflating prices and margins for incumbents like NVIDIA. JPMorgan implicitly benefits from this regulatory moat.
For blockchain, this is a critical nuance. Chinese mining pools control over 50% of Bitcoin hashrate. If export controls expand to include mining ASICs (as some US lawmakers have proposed), the supply of new hardware to Chinese miners could be restricted, further centralizing hashrate outside China or driving it underground. The strategist’s “2028” timeline is partly a function of this regulatory uncertainty—new fabs in the US and Europe are being built to reduce reliance on Taiwan, but they won’t be fully operational until 2028-2030. Until then, the supply chain is geopolitically fragile.
5. The Memory Component: HBM as a Bottleneck Within a Bottleneck
AI chips require high-bandwidth memory (HBM), which is itself supply-constrained. SK Hynix and Samsung are the only producers, and their HBM3e capacity is already sold out through 2025. The JPMorgan note did not dwell on this, but it is critical for blockchain: decentralized storage networks like Filecoin and Arweave use conventional NAND, not HBM. However, any blockchain that performs large-scale data processing (e.g., dYdX with on-chain order books) could benefit from faster memory. The HBM shortage means that the cost of high-performance computing will remain high, raising the barrier for computationally heavy DeFi applications.
Contrarian: What the Bulls Got Right—But Also Where They Miss
JPMorgan’s strategist is correct that AI demand is structural, not cyclical. The analogy to the 1990s internet infrastructure buildout is apt. Blockchain projects that require AI inference—such as decentralized AI agents or verifiable compute networks—will see growing demand, and the hardware shortage acts as a catalyst for price appreciation of the underlying tokens. Render and Akash, for example, have benefited from the narrative of GPU scarcity.
But the strategist underestimates two risks. First, the assumption that hyperscaler capex remains perpetually high ignores the possibility of a “AI winter” in spending if ROI disappoints. In 2022, Meta’s capex growth decelerated after a miss in advertising revenue. A similar phenomenon could occur if large language models fail to generate sufficient returns. Second, the analysis assumes that the current chip incumbents (NVIDIA, TSMC) will maintain their dominance. I am reminded of auditing the Bored Ape Yacht Club smart contract in 2021: the community believed the NFT utility was secure, but I found centralized metadata storage and gas inefficiencies. Similarly, the assumption that NVIDIA’s CUDA moat is unassailable may be challenged by open-source alternatives like PyTorch’s native support for AMD ROCm, or by custom ASICs from hyperscalers. The risk is that the “2028” bottleneck applies only to the current generation of chips. If a new, more efficient architecture emerges on a different node (e.g., chiplets on mature nodes), the bottleneck could shift.
Takeaway: Trust No One, Verify Everything—Including the Supply Chain
The JPMorgan note is a useful reality check for the blockchain industry. The next four years will be defined not by innovation in consensus or smart contract languages, but by the physical availability of silicon. Projects that depend on abundant, cheap compute must diversify their hardware dependencies or accept higher costs. I would advise blockchain builders to simulate their total cost of ownership under a prolonged shortage: assume GPU prices remain 30% above historical averages, assume lead times of 12 months for new hardware, and assume that the fastest chips are reserved for hyperscalers.
Final thought: The blockchain industry prides itself on trustless systems. But we have placed an enormous trust in a single company on a single island (TSMC in Taiwan) to deliver the computational backbone for decentralized AI and mining. Audit the code, not the pitch—but also audit the fab. Until we have geopolitically distributed chip manufacturing, the 2028 supply constraint will remain the unspoken risk behind every bull case for blockchain hardware.
"Sharding is easy; consensus is hard." So is building a chip fab. Let’s not confuse software agility with hardware reality.