Last March, in a rented room above a coffee shop in District 3, a developer from Da Nang asked me a question I could not answer.
We had spent three hours on node operation — bandwidth costs, disk pressure, the quiet economics of running your own verification instead of renting someone else's. Then he asked: if a chain can prove every transaction it has ever settled, why is the carbon accounting for the hardware that runs it still a PDF?
I thought about that question again this month, when Nvidia's latest sustainability disclosures crossed my desk. The company's supply-chain emissions — the category accountants call Scope 3 — have risen roughly 725% since fiscal 2020. The absolute figure now sits in the range of 15 to 16 million metric tons of CO2 equivalent per year, a total that places a chip designer owning no mines, no fabs, and no smokestacks somewhere in the neighborhood of a mid-sized Russian coal producer.
Nvidia has not concealed this. It is published, footnoted, and signed by officers who can be sued for misstating it. That is exactly what makes it worth examining. The most consequential number in the AI supply chain is one that no one outside a handful of corporate sustainability teams can independently verify. We have built the fastest computing infrastructure in human history on top of an accounting system that still runs on trust.
To see how a fabless company ends up with a carbon footprint larger than many heavy industrial firms, you have to understand how emissions are classified, and why the classification itself is the story.
Scope 1 covers direct emissions from owned operations. Scope 2 covers purchased electricity and heat. For Nvidia, both are modest relative to the whole — the company's own data centers, offices, and laboratories are close to a rounding error. Scope 3 covers everything else: purchased goods and services, upstream transportation and distribution, business travel, and the use of sold products. For semiconductor firms, Scope 3 is almost the entire total, typically 90% or more of what gets disclosed.
The physical chain runs roughly like this. Nvidia designs a die. TSMC fabricates it on a leading-edge node — N4 and N5 for Hopper, N3 and its variants for Blackwell and the parts that followed. Each wafer pass consumes enormous quantities of electricity, ultra-pure water, and a family of fluorinated process gases whose global warming potentials run into the tens of thousands relative to CO2. The die is then bonded to high-bandwidth memory stacks, sourced largely from SK Hynix, Micron, and Samsung, using advanced packaging — TSMC's CoWoS and its successors — a step that was for most of 2024 and 2025 the single tightest bottleneck in the entire AI supply chain. The finished module goes onto a board, the board into a server, the server into a rack.
TSMC is already among the largest single purchasers of electricity in Taiwan, taking a share of the island's power reported in the mid-single digits and rising. That number never appears in an Nvidia press release. It appears in Nvidia's Scope 3.
Power density is where the abstraction becomes physical. A single GB200 NVL72 rack draws on the order of 120 kilowatts. Ten thousand racks — a modest cluster by 2026 standards — pull more than a small city. And each generation raises the floor rather than lowering it, because performance gains now come as much from packing more silicon into the same volume as from shrinking transistors.
Then there is demand. The IEA has projected that data centers could consume close to 945 terawatt-hours annually by 2030, roughly double their 2024 level, with AI accelerators responsible for a disproportionate share of the growth. Those projections are contested in the details and uncontested in direction.
All of it sits on top of a market that has become the gravitational center of global equity. Nvidia's valuation pushed past four trillion dollars and brushed five, carried by hyperscaler capital expenditure that now exceeds the annual GDP of most nations. The emissions are not a side effect of that expansion. They are the expansion, expressed in a different unit. Every dollar of that capex has a mass, and the mass has a chemistry.
Nvidia's fiscal year also runs ahead of the calendar, which means the figures disclosed in any given report describe production decisions made twelve to eighteen months earlier. The emissions you can read about today reflect wafer starts ordered during the last great scramble for capacity. The next two reporting cycles, covering the Blackwell ramp and its successor, will very likely get worse before they get better — not because the company changed course, but because the physics of a supply chain arrives late.
Nvidia's stated position is that it targets net-zero across its value chain by 2040, that it matches its electricity use with renewable purchases, and that it works with suppliers on process efficiency. Those commitments are real, and they are also the easiest kind to make, because the hardest part of them — the supplier base — is the part the company does not control and, under current rules, does not have to prove.
The 725% figure deserves to be read slowly, because it is doing two jobs at once.
On one hand it is a growth story. Between fiscal 2020 and the most recent reporting year, Nvidia's data center revenue grew by roughly an order of magnitude. If unit volume rises tenfold and per-unit embodied carbon falls by a third, the total still rises more than sixfold. That is arithmetic, not negligence. The disclosure is not evidence that Nvidia became reckless. It is evidence that the world ordered an extraordinary quantity of silicon in a very short window, and that the supply chain had no slack to absorb it cleanly.
On the other hand, the pace of growth exposes how thin the measurement underneath really is. When a figure climbs that fast, the boundary of what is being counted almost always shifts with it. Suppliers get added to the inventory. Categories get reclassified between reporting years. Estimation methodologies get refined in ways that flatter the baseline. A company can appear to cut emissions by changing who counts as upstream, and nothing in the current framework prevents that. Nvidia is not uniquely exposed to this; every manufacturer with a deep supplier base is. What is unusual is the number of tiers involved and the number of jurisdictions those tiers span.
I spent part of 2017 doing forensic work on smart contract libraries, tracing reentrancy paths through multi-signature logic that could have drained hundreds of millions of dollars. The lesson I carried out of that year was not that code is fragile. It was that trust does not disappear when you remove a trusted party. It relocates. You trade a bank's promise for a compiler's correctness, and then you discover the compiler was written by people, reviewed by people, and deployed by people who were in a hurry.
Carbon accounting is the same problem wearing different clothes. The absence of a central authority does not mean the absence of a trust assumption. It means the assumption has been pushed down into a supply chain where almost nobody can see it. Tracing the code back to the conscience is the same motion as tracing an emissions figure back to the factory floor that produced it — you keep following until you reach a human being making a decision, and that is where the real accounting always lives.
This is the point at which the crypto industry usually raises a hand, and where I want to be careful rather than enthusiastic.
There is a genuine and growing category of protocols built around verifiable environmental data. The umbrella term is dMRV — digital measurement, reporting, and verification — and it spans satellite-linked forest monitoring, sensor networks that attest to industrial output, and cryptographic registries for land and resource claims. Toucan and KlimaDAO tried to bring retired carbon credits on-chain and build a liquid market around them. Regen Network has spent years on registry infrastructure. Open Forest Protocol and Silvi put field measurements on-chain with signed attestations. Akash, Render, and the various GPU-aggregation networks attack the compute side by renting out capacity that would otherwise sit idle.
The results have been mixed, and honesty requires saying so plainly. In 2023, a joint investigation found that a substantial share of tokenized credits traced back to projects whose additionality was questionable and whose baselines had been inflated. The tokens were, technically, immutable. The claims they carried were not. An on-chain record of an off-chain claim is only as strong as the oracle between them, and here the oracle was a human with an incentive to overstate. That is not a failure of cryptography. It is cryptography working exactly as designed on inputs that were never verified.
The more interesting frontier is not tokenized offsets. It is supply-chain attestation with cryptographic provenance. If a foundry, a memory vendor, and a packaging house were each to sign attestations about energy inputs, process gas consumption, and water use — and if those attestations were bound to specific lot identifiers — a downstream buyer would have something it could verify rather than trust. Confidential computing makes this plausible at the enterprise level, because a trusted execution environment can prove that a computation ran on real data without exposing the data itself. Zero-knowledge proofs can do the same for narrow statements: that a given wafer lot's energy intensity fell below a threshold, without disclosing the actual figure or the process recipe behind it.
None of this is deployed at scale in the semiconductor industry, and the reasons are mundane. Attestation infrastructure costs money. Suppliers guard process data the way they guard yields, because the data is the moat. And the buyers — a handful of hyperscalers — have little commercial incentive to expose their own upstream footprint to that degree of scrutiny when a summary table satisfies the disclosure requirement.
There is one more technical point that rarely makes it into the coverage. The embodied carbon of an accelerator is amortized across its service life, and that service life is contracting. A100 clusters deployed in 2020 are being retired. H100 fleets are being displaced by Blackwell, and Blackwell will be displaced in turn by whatever comes next. Each refresh front-loads a new tranche of fabrication emissions and produces a new stream of electronic waste. The industry has quietly moved from something like a five-year replacement cadence toward two or three. Efficiency gains per floating-point operation are real and large. Efficiency gains per dollar of deployed capital are much weaker than the marketing implies, because the denominator keeps getting replaced before it can pay back its own manufacturing cost.
Listening to the silence between the blocks is the whole discipline here. The gaps in a ledger tell you more than the entries do. What Nvidia does not disclose — the allocation of emissions across product lines, the carbon intensity of individual supplier tiers, the amortization assumptions behind use-phase accounting — is where the analytical value sits. The published total is a headline. The unpublished decomposition is the argument.
And the rules governing that decomposition are being rewritten right now. The GHG Protocol is under revision. The Science Based Targets initiative has been pulled in opposing directions by companies that want the boundary loosened. The EU's corporate sustainability reporting directive is being pared back under industry pressure. Governance is not a vote; it is a vigil — and at the moment the vigil is being kept by a shrinking number of people with no enforcement power and no subpoena.
The reflex comparison — that Nvidia's footprint now rivals a coal producer — is emotionally satisfying and analytically close to useless.
A coal producer's emissions are combustion. They occur once, they occur fast, and they stop when the plant closes. The carbon embodied in an accelerator is capital: spent at fabrication, amortized across years of productive service, much of it powering medical imaging, weather simulation, protein folding, and materials discovery. Comparing a stock of embodied industrial carbon to a flow of combustion carbon flatters the headline and muddies the policy question. The right comparison is not Nvidia against coal. It is the AI buildout against every other use of the same electricity, the same water, and the same leading-edge wafer capacity.
That is a harder case to make, and a more honest one. It also admits unwelcome possibilities. If marginal AI compute delivered meaningful decarbonization — better grid forecasting, better catalyst discovery, better logistics — the trade could be worth defending. I have seen that argument made confidently many times and verified rigorously almost never. It is precisely the kind of claim that attestation infrastructure would settle, and precisely the kind nobody is building, because a negative finding would be commercially inconvenient.
The deeper blind spot, though, is not Nvidia's. It is ours.
Between 2021 and 2023, this industry endured years of hostile inquiry about energy consumption. I spent much of that period defending the legitimacy of proof-of-work to people who had already decided the answer. Then Ethereum merged and cut its electricity draw by more than 99%. Bitcoin's consumption kept climbing, but at a rate that now looks almost quaint beside what the AI buildout has done. And the same legislators, the same ESG funds, and the same journalists who once demanded that every validator justify its kilowatt-hours have gone conspicuously quiet.
I do not think that silence is a conspiracy. I think it is a value judgment about which industries deserve scrutiny. Compute for intelligence is treated as progress; compute for money is treated as waste. That distinction may even be defensible. But it should be argued in public, not assumed in private, and the crypto community should stop pretending the asymmetry is principled rather than convenient.
There is a second blind spot closer to home. Our answer to the emissions question cannot be "put it on-chain." Decentralized compute networks do not reduce the carbon intensity of a wafer fab. They redistribute who rents the output. DePIN protocols that aggregate idle GPUs are useful and I am glad they exist, but they are not a climate strategy, and claiming otherwise is the same category error as claiming that tokenizing a credit made a forest grow.
Truth is the only immutable asset. Everything else — including the token minted to represent it — can be reorganized, reclassified, or quietly retired.
So what do we do with a number nobody can verify?
The honest answer is that we stop treating disclosure as verification and start treating verification as infrastructure. That means buying attestation the way we buy redundancy: boringly, unglamorously, and before we need it. It means demanding that Scope 3 data carry cryptographic provenance the same way we demand that a transaction carry a valid signature. It means accepting that the tools we built for money may be more valuable pointed at material flows than at speculative ones.
For those of us watching a sideways market and looking for signals, this is a slow one rather than a fast one. It will not move a price this week. It will decide which infrastructure narratives survive the next regulatory cycle, and which get quietly written down alongside the credits that never should have been minted.
I keep returning to the developer in District 3. His question deserved a better answer than I gave him that afternoon. A chain can prove its own history. It cannot yet prove the history of the machine it runs on.
We build bridges from the ashes of belief. The next one has to reach from a block header to a wafer lot.