The Silicon Underneath the Token: What Broadcom's 86% Quarter Really Reveals About Decentralized Compute

Over the past ninety days, a single fabless chip designer captured more of the AI value chain than every decentralized compute network on earth combined — and most crypto founders never noticed. Broadcom's AI revenue grew 86% year over year. Read that slowly. We didn't build the compute layer of the next internet. We rented it from a company most people in this industry cannot name, wrapped it in a governance token, and told our communities we were decentralizing the future. Then the landlord sent a new invoice.

I want to be precise about what that 86% actually measures, because the number as reported is doing something slippery. Depending on how you read the source — and I have read it four times, coffee going cold beside me — it is either a company-wide revenue surge or an AI-segment surge. Those are very different claims. One describes a diversified infrastructure giant riding a wave. The other describes a company whose entire valuation now rests on a single, violently cyclical customer category. In a bear market, that distinction is the difference between a hedge and a trap. So before I tell you what this means for crypto, let me be honest: the original brief was thin. One hard data point, ambiguous in scope. Everything else I am about to give you is me doing what I always do when the tape is quiet — pulling in the industry baseline, the packaging bottleneck, the export-control map, and the on-chain signals that the headline number is trying to hide.

The reason this matters to you, whether you hold DePIN tokens, run a validator, or just watch hashrate charts at 2 a.m., is simple. Crypto's most ambitious current narrative — that compute itself can be decentralized, permissionless, and owned by its users — runs on top of the exact supply chain Broadcom just monetized. Every AI inference job you route to a tokenized GPU network, every proving cost you pay on a ZK rollup, every mining farm you underwrite, terminates in the same place: a handful of fabs, a single dominant packaging line, and a photolithography monopoly in the Netherlands. That is the substrate. And the substrate is not decentralized. It never was.

I have been watching this layer for nineteen years, first as a junior consultant auditing fiat systems in Chicago, then as someone who abandoned a scheduled audit to spend three months building a crude proof-of-knowledge demo after a late-night reading of the ZK-SNARK papers. That detour taught me something I have never stopped repeating to founders: the math can be trustless, but the machines that run the math are owned by someone. The whole dream of verifiable computation collapses if the hardware that produces the proof is a single point of failure. So let me walk you through what the Broadcom number actually sits on top of, and why the crypto industry is misreading it in a way that will cost people real money this cycle.

The first thing to understand is that the AI silicon boom is not a rising tide — it is a migration. Broadcom does not sell chips the way NVIDIA sells chips. Broadcom sells custom silicon, application-specific integrated circuits, built to a hyperscaler's own specifications. Google's TPU. Meta's training and inference accelerators. ByteDance's internal compute. These are not general-purpose products you buy off a shelf; they are bespoke engines, designed by the customer, manufactured by Broadcom, fabbed by TSMC. The 86% growth is not a company riding a generic AI wave. It is the measurable speed at which the largest compute buyers on the planet are defecting from the general-purpose GPU to their own private silicon.

That migration is the single most important structural fact in the compute economy right now, and the crypto industry has almost entirely failed to price it. When I forked three AMM protocols during the 2020 DeFi Summer — not to farm yield but to study how governance actually behaves under stress — I learned that the most dangerous moment for any system is not when it fails, but when it succeeds in a way that centralizes power while everyone is celebrating. The AI ASIC migration is exactly that moment for compute. The hyperscalers are succeeding so spectacularly at custom silicon that they are quietly removing themselves from the open market for compute altogether. And the open market for compute is where crypto's entire DePIN thesis lives.

Think about what a decentralized physical infrastructure network actually promises. It says: instead of one giant data center owned by one giant company, we will pool thousands of independent machines — GPUs in basements, rigs in warehouses, idle cards in gaming PCs — and coordinate them with a token so that anyone can buy and sell compute without permission. It is a beautiful idea. I have written about it with genuine hope. But the Broadcom number tells you who is winning the actual compute war, and it is not the pool of independent machines. It is the hyperscaler with a bespoke ASIC that is thirty to fifty percent cheaper per unit of work than the GPU the DePIN network is trying to aggregate. When your competitor's cost structure is half of yours, decentralization stops being a feature and starts being a tax.

Liquidity isn't the lifeblood of a market — it is the memory of who showed up last time. The same is true of compute. The compute that matters is not the compute that exists; it is the compute that reliably shows up, on time, at a predictable cost, for the workloads that pay the most. And right now, the workloads that pay the most — frontier model training, large-scale inference — are showing up at Broadcom and TSMC, not at your tokenized GPU marketplace. The DePIN networks are real, and some of them are well-built, but they are competing for the scraps of the inference market while the feast is served behind a fab's cleanroom door.

Here is where I have to bring in the packaging bottleneck, because it is the hidden variable almost nobody in crypto talks about and it determines everything. Broadcom's large AI ASICs do not just need wafers. They need advanced packaging — specifically TSMC's CoWoS, chip-on-wafer-on-substrate, a 2.5D integration technology that stacks the compute die next to high-bandwidth memory and ties it together with silicon interposers. The real constraint on AI chip shipments in 2024 and 2025 was never wafer capacity. It was CoWoS capacity. And CoWoS capacity is allocated, not auctioned. The biggest customer gets first call, and the biggest customer is NVIDIA.

Sit with that for a moment. A decentralized compute network's ability to source hardware, in the end, depends on where it falls in a queue managed by a single Taiwanese packaging line, prioritized by a single American chip designer, for a single dominant customer. There is no token that fixes that. There is no governance vote that moves you up the queue. The decentralization of compute is bottlenecked not by cryptography but by a physical packaging step that no amount of on-chain coordination can replicate. This is the kind of thing I mean when I say the substrate is not decentralized. You can decentralize the coordination layer. You cannot decentralize the cleanroom.

I learned a version of this lesson the hard way. When I co-founded Artory in 2021 — a project trying to link NFT ownership to real-world reputation — I assumed the hard part would be the cryptography. It wasn't. The hard part was the physical and legal substrate: verifying that a human had actually done the volunteer hours, that the record was real, that the reputation meant something off-chain. The blockchain was the easy, elegant part. Everything underneath it was a mess of trust. DePIN has the same shape. The token is elegant. The silicon underneath is a mess of monopolies.

This is where the bear market changes the entire calculation. In a bull market, you can ignore cost structure because liquidity hides inefficiency. In a bear market, cost structure is the only thing that survives. And the cost structure of decentralized compute is brutal. A tokenized GPU network has to pay for hardware, power, bandwidth, and the coordination overhead of matching thousands of independent operators against unpredictable demand. A hyperscaler running bespoke ASICs pays for hardware, power, and bandwidth — and its hardware is cheaper, its utilization is higher, and its demand is contractually locked in years in advance. The DePIN network is running a business model that only works when the centralized alternative is capacity-constrained. And right now, the centralized alternative is capacity-constrained on packaging, not on compute — which means the moment CoWoS capacity expands, the centralized players can flood the market and crush the decentralized price floor.

So which protocols are bleeding? Look at the ones whose token emissions are subsidizing hardware that has no independent economic reason to run. If a GPU network's operators only stay online because the token reward exceeds the cost of electricity, you are not looking at an infrastructure business. You are looking at a ponzi with a cooling fan. The Broadcom number tells you the real compute economy is profitable without subsidy. That is the bar. If a DePIN network cannot clear that bar, it is not infrastructure. It is a bet on future demand that may never arrive.

Now let me bring in the part that should genuinely worry anyone who cares about crypto's values, because it is the part the original brief completely missed. The semiconductor industry is splitting into two geopolitical camps, and Broadcom and Intel sit on opposite sides of that split. Broadcom is a globalization winner: it depends on TSMC in Taiwan for manufacturing, on CoWoS packaging, and on a global customer base that includes Chinese hyperscalers like ByteDance. Intel is a localization winner: it is the single largest beneficiary of the American CHIPS Act, it is building fabs in Arizona and Ohio and Germany, and its strategic value is precisely that it manufactures on American soil.

This means the geopolitics of compute are the exact inverse of the geopolitics of crypto's stated values. Crypto believes in borderlessness — permissionless access, global liquidity, no gatekeepers. The compute layer is racing in the opposite direction: toward export controls, toward sovereign fabs, toward a world where the physical means of production are nationalized in everything but name. When the United States restricts what can be sold to whom, and when China retaliates with gallium and germanium export controls, the compute substrate fragments. And a fragmented substrate is hostile to the borderless coordination layer crypto is trying to build on top of it. We didn't build a global compute commons. We built a token that depends on a supply chain being actively weaponized.

I have watched this dynamic up close. In 2025 I collaborated with a Chicago-based AI ethics lab to draft an ethical constraint protocol for autonomous DAO treasuries — a document that tried to combine legal theory with smart contract logic so that an AI agent managing a multi-sig wallet could never move funds outside human-approved boundaries. What struck me during those months was that every ethical safeguard we designed assumed a stable, accessible compute base. If the compute base fragments — if a jurisdiction can simply deny your network the silicon it needs — then your ethical constraints are academic. You cannot govern an autonomous treasury that cannot run. The hardware layer is the precondition for every value the crypto industry claims to hold.

And this brings me to Intel, because Intel's story is the mirror image of Broadcom's, and the crypto industry keeps conflating them. The brief called Intel's situation a "recovery." I want to interrogate that word, because a recovery can be technical or it can be financial, and the two are not the same thing. Intel's technical recovery hinges on one process node: 18A, its RibbonFET gate-all-around transistor combined with PowerVia backside power delivery. If 18A yields well enough to attract external foundry customers, Intel rejoins the leading edge alongside TSMC's 2nm and Samsung's SF2. If it doesn't, Intel is a company that spent hundreds of billions of dollars on fabs it cannot fill.

Here is the trap in the "recovery" narrative. Intel's capital expenditure has historically run at thirty to forty percent of revenue — an enormous, structural drag. Every one of those fabs carries depreciation that crushes gross margin for years, whether or not the fab produces anything anyone buys. Intel's gross margin has collapsed from the sixty-percent range toward the thirty-to-forty range, and the reason is not that Intel forgot how to sell chips. It is that Intel is carrying the depreciation of a manufacturing base that is underutilized and technologically behind. A financial recovery — cost cuts, layoffs, asset sales — looks identical to a technical recovery in a single quarterly headline, but only one of them is durable. If Intel's improvement comes from headcount reduction and divestitures rather than from 18A yielding, then the recovery is cosmetic. The crypto industry, which loves a turnaround narrative, is especially vulnerable to mistaking the two.

I have a rule from the 2022 crash that I have never abandoned. When my portfolio collapsed and my mood went with it, I forced myself to stop looking at price and start looking at code activity — the silent builders who kept shipping while everyone else capitulated. That discipline is why I trust engineering output over narrative. So when I evaluate Intel's recovery, I do not look at the press release. I look at whether external customers are signing 18A tape-outs. And when I evaluate a DePIN network, I do not look at the token price. I look at whether operators stay online without emissions. The Broadcom number passes that test on its face: the demand is real, contractual, and paid for. Most DePIN tokens do not. That is the survival question of this bear market, and it is the only question that matters.

Let me now push into the contrarian angle, because I have spent most of this piece setting up a case that decentralized compute is structurally disadvantaged, and I want to complicate my own argument. The easy conclusion is that DePIN is doomed, that the centralized silicon stack wins, that the token is a fantasy. I do not believe that. Here is why.

Decentralization was never going to happen at the hardware layer, and the mistake is expecting it to. The hardware layer has always been centralized — in mainframes, in fabs, in the steam engines of every prior industrial revolution. The value of crypto's coordination layer is not that it decentralizes production; it is that it decentralizes access, pricing, and ownership of the surplus. Broadcom can own the fab and the packaging line and the customer contract, and a decentralized network can still own the marketplace where the resulting compute is priced and sold — if, and only if, it can offer something the centralized market structurally cannot. And there is one thing it can offer: verifiable, permissionless access for workloads that the centralized market refuses to serve. Small inference jobs. Censorship-resistant compute. Idle capacity in jurisdictions the hyperscalers ignore. These are not glamorous, and they will never produce an 86% growth headline. But they are real, and they are defensible, precisely because the centralized players have no incentive to serve them.

This is the same lesson I learned running governance jams during the 2020 DeFi Summer. We forked three AMMs and organized weekly sessions that pulled in five hundred participants, and the thing that actually moved turnout — forty percent in a single quarter — was not better tokenomics. It was that the community felt ownership of a market the incumbents were not serving. Decentralization works at the margin, not at the center. It captures what the center discards. Anyone promising that a tokenized GPU pool will out-train OpenAI on bespoke silicon is selling you a story that the Broadcom number directly refutes.

And here is the second complication. The centralized compute stack is not stable either. Broadcom's 86% growth comes with a specific, documented risk: extreme customer concentration. Its AI business depends on a handful of hyperscalers, and those hyperscalers have their own chip teams. The same dynamic that makes Broadcom powerful — being the supplier to a few giant customers — makes it fragile, because those customers can and will bring design in-house, or defect to a competitor. A supplier whose fate rests on three customers is not a moat; it is a hostage situation with better margins. The original brief flagged this as a risk, and it was right to. But it under-read the implication. If Broadcom's concentration is a risk, then the whole compute economy is concentrated, which means the centralized alternative to DePIN is itself a single point of failure. That is precisely the argument for a decentralized fallback — not because it is more efficient, but because it is more resilient. Resilience is what you buy in a bear market. Efficiency is what you buy in a bull market.

There is a third complication, and it is the one I find most interesting as someone who has spent years on proving systems. ZK rollups — the technology I have evangelized since that first late-night reading of the SNARK papers — have a compute cost problem that is structurally identical to the AI compute problem. Proving is expensive. Generating a validity proof for a complex state transition consumes enormous compute, and unless the price of that compute falls dramatically, the economics of running a prover are brutal. Operators are bleeding money on proving costs, and the only thing that saves them is cheap gas on the settlement layer — which, in a bear market, is exactly the thing that evaporates. ZK rollups and AI networks share the same hidden dependency: they are both, at bottom, compute businesses pretending to be software businesses. And compute businesses live or die on the cost of silicon, power, and packaging — the very things Broadcom just proved are the real profit pools.

I will say the uncomfortable part plainly, because I have earned the right to after nineteen years. We told ourselves that blockchains would disintermediate the compute economy. They didn't. They inserted a new intermediary — a token — on top of a supply chain that was already captured. The Lightning Network spent seven years promising cheap, fast, decentralized payments and never solved the routing and channel-management complexity that keeps it a niche curiosity; the same pattern of elegant coordination layered on intractable physical and economic substrate shows up everywhere in this industry, from payment channels to GPU marketplaces. And on the programmable-liquidity side, we celebrated Uniswap V4's hooks as the moment the DEX became composable Lego — but the complexity spike is quietly scaring away the very developers the feature was meant to attract, the same way a DePIN SDK that takes three weeks to integrate scares away the operators who would make the network real. Programmable everything is not the same as usable anything.

So what is the actual, defensible thesis for compute in crypto, stripped of the evangelism? It is this. The centralized compute stack will keep winning the frontier — the training runs, the big inference clusters, the workloads that justify a bespoke ASIC. Broadcom will keep printing money as long as the hyperscalers keep defecting from GPUs, and that migration has years left to run. But the centralized stack has three structural fragilities that crypto's coordination layer can exploit at the margin: it is geographically concentrated and therefore geopolitically exposed; it is customer-concentrated and therefore commercially exposed; and it is capital-intensive and therefore cyclical. Every one of those fragilities is an opening for a decentralized network that does not try to compete on the frontier, but instead owns the long tail of compute that the center cannot profitably serve.

The winners in this bear market will not be the networks with the biggest token emissions or the loudest decentralization claims. They will be the networks whose operators stay online when the subsidy ends. They will be the networks that price compute against a real cost of silicon and power, not against a token that only goes up. They will be the networks that treat the hardware layer honestly — as centralized, constrained, and geopolitical — and design their coordination layer around that reality instead of pretending it away. Freedom isn't the absence of dependency; it's the presence of consent, and the honest DePIN network is the one that tells its community exactly who owns the silicon underneath their tokens.

I keep coming back to the CHIPS Act, because it is the tell that almost everyone missed. Intel's recovery, to the extent it is real, is heavily policy-dependent — roughly eight and a half billion dollars in subsidies and loans, plus the strategic value of being the American foundry. Strip out the subsidy and the recovery's quality drops sharply. That should tell you something about the compute economy as a whole: it is not a pure market. It is a market shaped by national strategy, export controls, and the fear of a single point of failure in a single strait. Crypto, which prides itself on being immune to state power, is building its most ambitious infrastructure narrative on top of the most state-shaped industry on earth. That is not a contradiction to be ashamed of. It is a reality to be designed around.

And the geography is only going to harden. China's export controls on gallium, germanium, and antimony will not cripple Broadcom or Intel directly — those materials matter most to mature-node and compound semiconductors — but they signal a long-term decoupling that will raise costs across the board. Every layer of the stack, from lithography in the Netherlands to packaging in Taiwan, is being pulled into a national-security frame. For crypto, this means one thing above all: the days of assuming a frictionless, global compute substrate are over. The protocols that survive the next cycle will be the ones that treat compute as a scarce, contested, geographically embedded resource — because that is what it actually is.

I want to end where I started, with the number, because the number is the whole argument compressed into two digits. Eighty-six percent. That is the rate at which a single fabless designer grew its AI revenue while the decentralized compute sector, in aggregate, struggled to hold its valuation. If you are holding a DePIN token, or building a compute network, or writing the next whitepaper about decentralizing the data center, you need to sit with that gap and decide what it means. It does not mean decentralization is dead. It means decentralization was never going to win at the layer where physics and capital and nation-states decide the outcome. It was always going to win at the layer where coordination, access, and ownership can actually be changed — the layer crypto is genuinely good at.

The question for the next cycle is not whether we can decentralize silicon. We can't, and pretending otherwise has cost this industry years and billions. The question is whether we can build a coordination layer so useful, so resilient, and so honest about its dependencies that the centralized stack has no choice but to route through it. That is a harder dream than the one we sold in 2021. It is also the only one that survives contact with a cleanroom. And in a bear market, survival is the entire game.

So watch the silicon. Watch the packaging queue. Watch who signs the next 18A tape-out and who takes the next CoWoS allocation. Because that is where the future of compute is actually being decided — and the tokens will only ever be a footnote to what the fabs already decided.