I still keep a GitHub issue pinned in a browser tab I never close. It is from 2017, a thread where a twenty-five-year-old version of me, sitting in a Beijing apartment at two in the morning, walked through twelve logic flaws in a multi-signature wallet implementation. I submitted it not for a bounty but because I could not sleep knowing that early adopters were about to trust a contract with centralized failure points baked into its escape hatches. Nine years later I reread that thread whenever the market confuses a headline for a thesis. This week gave me a reason to reread it again.
The headline, as it reached me, was that DeepSeek had deployed its models "at scale" on Huawei's Ascend chips inside a data hub in Inner Mongolia. That was the entirety of the claim. No chip count. No commissioning date. No distinction between a training run and a serving endpoint. No primary source โ only a crypto vertical relaying what was most likely a secondhand digest of a report that was itself probably secondhand. And within a single news cycle, a hundred threads declared the collapse of American AI dominance, the maturation of a sovereign compute stack, and โ this is the line that made me close the laptop and put the kettle on โ the vindication of decentralization.
Follow the fear, not the chart. The chart in my feed was euphoria. The fear underneath it was something I recognize, because it lives in me too: the fear that the future is being built somewhere else, by people who do not share our vocabulary, and that we will mistake their concrete for our ideology.
Here is what the charts won't tell you. The story is real. It matters. And almost everything the market extracted from it is wrong โ not because the facts are false, but because the frame is borrowed. We are reading industrial policy as though it were a story about us.
What We Actually Know
Let me be disciplined about the evidence, because if you can't be disciplined about evidence in a bull market, you are not an analyst, you are a mirror. The original item contains, at most, two factual assertions, and they are near-tautologies of each other: DeepSeek is deploying on Huawei Ascend; the deployment is in Inner Mongolia. Everything else โ the scale, the strategic meaning, the geopolitical weight โ is the writer's interpretation layered on top of a press release we never see.
So let me separate the layers. What is almost certainly true: DeepSeek has been moving toward domestic silicon for over a year, driven by forces it did not choose. What is likely true: the deployment is primarily inference, with training still dependent on legacy Nvidia inventory. What is unknown and load-bearing: the chip model, the cluster size, the interconnect topology, the commissioning date, the financing structure, the operating entity, and โ most importantly โ whether any training at all is happening on those racks.
That last distinction is not a technicality. It is the entire story. And the fact that no one in my feed paused to ask it tells you more about the crypto industry's reading habits than about DeepSeek's engineering.
The Distinction That Decides Everything
I want to explain why "training versus inference" is not a footnote, because I have spent enough nights inside software stacks to know that the difference between the two is the difference between a marketing claim and an operational reality.
Inference โ serving a trained model to answer queries โ is a comparatively forgiving workload. A model like DeepSeek's flagship, a mixture-of-experts architecture with a very large total parameter count but a much smaller active parameter count per token, activates only a fraction of its weights on any given forward pass. That sparsity is a gift to a young software stack. You need a working set of operators, a reliable memory hierarchy, and enough bandwidth to feed the active experts. You do not need the entire training graph to be numerically stable across hundreds of billions of parameters. You do not need gradient synchronization at the collective-communication layer to behave. You are running the machine in one direction, at steady state, over and over.
Training is the opposite of forgiving. It requires every operator in the forward and backward pass to be covered, not a curated subset. It requires the collective-communication primitives โ the all-reduce and all-gather calls that synchronize gradients across thousands of chips โ to be not merely present but production-grade in both correctness and throughput. It requires mixed-precision arithmetic to remain stable across weeks of iteration, because a loss spike in week three costs you the run, not an afternoon. It requires checkpoint-and-resume to work flawlessly, because at large scale hardware failure is a scheduled event, not an emergency.
When I audited that multi-sig in 2017, I learned a version of this lesson at a much smaller scale. A wallet can look correct in a hundred happy-path tests and still leak value the moment an adversary sequences transactions in a way the developer never modeled. The happy path is not the system. The failure path is the system. And on training clusters, the failure path is where the entire software maturity of a hardware ecosystem gets exposed.
This is why the migration from CUDA to Huawei's CANN stack is asymmetric between the two workloads, and why the asymmetry is roughly an order of magnitude. Moving an inference workload from one to the other is engineering. Moving a training workload is re-engineering, and sometimes re-researching, because the numerics do not always survive translation.
The Evidence in the Weights
Here is a piece of technical detective work that almost nobody in the thread did, and if you can, you should train yourself to do it, because it is where the information gain lives.
DeepSeek's own releases carry a stealth fingerprint of its hardware intentions. Its intermediate model versions adopted an FP8 training and inference format โ an eight-bit floating-point representation with specific exponent and mantissa semantics โ that lines up, with suspicious precision, against Huawei's announced roadmap for its upcoming Ascend generation, which is slated to include native support for the same family of low-precision formats. That alignment does not happen by accident. When a model architecture and a chip roadmap share a numerical vocabulary this specific, the collaboration is not a purchase order. It is a co-design relationship with a lead time measured in quarters, sometimes years.
Which is to say: the Inner Mongolia headline is not the beginning of the story. It is the surface of a story that began long before any journalist noticed the concrete.
But the same evidence cuts the other way, and this is the part the euphoria skipped. It is one thing to align formats on a roadmap. It is another to run a frontier-scale training job on first-generation production silicon for those formats. There are public reports, consistent with what I hear from engineers who would know, that DeepSeek's attempts to train on Ascend hardware at the frontier encountered stability and performance problems significant enough to delay a release cycle. I cannot verify the specifics. But the direction of the signal โ training on the new stack is hard, inference on it is plausible โ matches everything the industry knows about how compute ecosystems mature. Inference always goes first. Training follows only when the toolchain stops surprising you.
So when someone tells you DeepSeek has "deployed on Huawei," the honest question is not whether they did. The honest question is what, exactly, they deployed โ and the most likely answer, the one the evidence supports, is a serving cluster that takes pressure off a training fleet still running on the older hardware they bought before the export controls closed.
Why Huawei Wants This More Than DeepSeek Does
Follow the incentives, not the press release. DeepSeek did not wake up one morning and decide to be a nationalist. It woke up in a market where its only legal option for new compute was domestic. The decision was made for it. In that light, the deployment is less a triumph than a consequence โ the result of a supply line being cut, not a preference being revealed.
Huawei, by contrast, chose this. And the choice is revelatory.
The hardest problem for Ascend was never peak arithmetic. Peak numbers are marketing; they move. The hard problem was software: operator coverage, debugging tooling, the thousand small frictions that make a stack pleasant or intolerable to the engineers who live inside it. A vendor can write all the whitepapers it likes about theoretical throughput. None of it substitutes for a reference workload โ a famous, respected model, running at scale, in production, with users, observed by skeptical third parties.
DeepSeek's flagship is close to the perfect stress test. It combines a custom attention mechanism with fine-grained mixture-of-experts routing at low precision. That is exactly the kind of workload that punishes a young software stack: unusual operators, irregular memory access, tight tolerances on numerics. If Huawei can make that model run well, the sales conversation for every subsequent customer changes. If it cannot, no amount of marketing recovers.
This is why I read the event as an unequal exchange. Huawei borrows DeepSeek's reputation to buy credibility for its ecosystem. DeepSeek pays for that credibility with efficiency losses, engineering distraction, and technical risk. It is a collaboration, but it is not a symmetric one โ and given the policy environment, DeepSeek's room to walk away is thin. The brand of the world's most respected open model, spent to validate a domestic chip stack, is a transaction whose full price will not be visible for two or three quarters.
The Bottleneck Nobody Names
If you want the one physical fact that governs whether this deployment becomes the first of thousands or the last of a handful, it is not in the datacenter. It is in the memory.
Every credible accounting of Ascend's scaling problem points to a supply constraint on high-bandwidth memory โ the stacked memory that sits beside the accelerator and feeds it at absurd bandwidth. That memory has, for years, been supplied predominantly by a small number of overseas vendors operating under the same export regime that pushed DeepSeek toward Ascend in the first place. Domestic alternatives exist and are progressing, but the yield and cost curves of a first-generation competitor to a mature HBM stack are not something you assert into existence on a roadmap slide.
So the beautiful, brutal question is this: can the deployment be supplied? A stated intention to install a large cluster is not an installed large cluster. If the memory cannot be sourced at volume, the "at scale" adjective quietly becomes a smaller number, and the demonstration effect โ the whole reason Huawei wanted this โ evaporates along with it.
I keep thinking about the multisig audit here too. The flaw that mattered was never the one in the headline; it was the one three layers down, in the assumption nobody checked because it had always held. For Ascend, the assumption that has always held is that memory will arrive. That is the assumption to watch, not the accelerator's peak number.
The Economics They Did Not Model
Now let me do the thing I actually trained for and talk about money, because a bull market is exactly when people stop asking whether a deployment is economical and start asking only whether it is ideological.
DeepSeek's defining commercial strategy is aggressive pricing. Its API sits at or near the bottom of the global price table for frontier-class capability, and its reasoning-tier models undercut closed competitors by wide margins. That pricing is only sustainable if the cost per token served is low. Margin in the API business is thin by design; the whole model is a volume bet on efficiency.
Which means every percentage point of inference throughput lost to a less mature software stack is not an abstraction. It is a direct subtraction from a margin that was already thin. If the new hardware serves meaningfully fewer tokens per second per watt than the older hardware it replaces, DeepSeek faces an impossible trio: absorb the loss, raise prices and lose the price leadership that is its identity, or subsidize the transition from reserves.
There is one factor that makes the third option liveable, and it is easy to miss. DeepSeek is backed by an in-house quantitative fund, which means it is not dependent on external capital, has no quarterly earnings pressure, and can absorb the friction of a hardware transition that a venture-funded competitor could not. That is not a small detail โ it is the structural reason this exact company can do this exact thing. The same move inside a younger, more competitive lab would be a financial event. Here it is a strategic cost, quietly booked.
And the countervailing factor โ the one that makes the economics less bleak than they look โ is the location. This is where the story stops being purely about chips and starts being about geography, which is another thing the thread completely ignored.
Power, Water, and the Quiet Subsidy
A datacenter is, at bottom, a machine for converting electricity into computation while trying not to cook itself. Every inefficiency in that conversion shows up on the invoice.
Inner Mongolia is not an accident. Its annual average temperature sits in the single digits Celsius, which means for much of the year the cooling burden can be met by pulling cold air through the hall rather than by running mechanical chillers. That pushes the facility's power-usage effectiveness toward the low end, near 1.2, which is the ratio of total facility power to the power that actually reaches the chips. In a hot, crowded coastal city, the same workload might carry a cooling overhead twice that, or more.
The second gift is the price of the electricity itself. Western and northern regions of the country sell power to industry at a discount measured in tens of percent against the coastal east โ the delta has been shrinking as renewables proliferate, but it remains real. For a cluster whose lifetime electricity bill is a figure with seven or eight digits, that discount compounds year after year.
The third gift is policy. The national "data centers in the west, computation demands in the east" framework directs computation-hungry workloads toward the interior precisely because the power and the land are there and the population is not. A cluster placed in Inner Mongolia satisfies two official narratives at once โ domestic chips for sovereignty, green-leaning western hubs for sustainability โ and that double satisfaction is worth more to the operating entity than any single electricity tariff. Facilities in these hubs have enjoyed access to computation vouchers, preferential land, and utility arrangements that never appear on a public balance sheet. The economics of a deployment like this are not legible from the outside because the most important terms are not published.
I want to be precise about what this does and does not solve. Cheaper power is a real and durable advantage. But it does not repair a throughput deficit in the software stack; it only softens the blow. If the tokens served per second per watt are lower, low electricity prices reduce the loss without eliminating it. Power is a discount on the mistake, not a cure for the mistake. Anyone who claims Inner Mongolia makes the economics unambiguously positive is doing arithmetic with one hand tied behind their back โ they are counting the savings and ignoring the throughput.
The Long-Distance Problem They Bury
There is a further cost that the euphoria never touches, and it is temporal rather than financial.
Inner Mongolia is far from the users. For a training job, that does not matter โ nobody is waiting for a gradient to come back. For batch inference, it barely matters. For the interactive, latency-sensitive services that increasingly define where AI revenue lives, distance translates into milliseconds, and milliseconds translate into experience. The round-trip between a western hub and a coastal user adds a delay that, on its own, is tolerable; stacked on top of a model that is already not the fastest to first token, it becomes a designed-in disadvantage.
Which tells us something structural about what this cluster is actually for. The geography nudges it toward the workloads that do not care about distance: training, fine-tuning, batch processing, offline evaluation. That is a quiet argument โ not proof, but an argument โ that whatever is happening on those racks is closer to the heavy, latency-indifferent side of the house than to the customer-facing side. And that, in turn, is one of the few pieces of circumstantial evidence pointing against the simplest reading that this is all just inference. The signal is faint. I will not pretend it is strong. But when the headlines are loud, faint signals are where the truth hides.
The Crypto Blind Spot
Now the part I actually came to write, the part that has nothing to do with Huawei and everything to do with us.
Watch how my own feed consumed this story. An industry that has spent a decade insisting that centralized infrastructure is the enemy, that credible neutrality is the only thing worth building, that the whole point of the last fifteen years was to escape dependence on a handful of gatekeepers โ that industry read a story about two enormous institutions building a tightly coupled, jurisdictionally confined, geopolitically motivated compute stack, and cheered, because the institutions were not American.
That is the blind spot. We have confused the enemy's name with the enemy's shape.
Centralization is not a nationality. It is a topology. A data center in Inner Mongolia running a state-adjacent chip and a nationally championed model is not more decentralized than a data center in Virginia running the same class of hardware; it is differently owned. Ownership matters, of course โ it always does โ but it is not the same axis as distribution, and we have spent so long fighting on one axis that we have forgotten the other exists.
Here is the discipline I would ask of anyone who cares about this space. When a piece of news arrives, ask not whether it favors our side, but whether it changes the topology. Does it move compute toward many independent operators or fewer? Does it increase the number of parties who can verify a claim, or decrease it? Does it make a single point of failure more or less survivable? Ask those questions and the Inner Mongolia story reads very differently. A triumph of sovereign capacity can also be a data point for concentration. Both things are true. The thread-boys could only hold one.
The Verifiability Problem, Restated
This is where my current work lives, so let me be honest about my bias โ and then make the argument anyway, because it survives the bias.
If AI capability is going to be concentrated in a handful of very large compute installations, then the question that matters to a decentralized world is no longer who owns the chips. It is who can verify what was done with them. The training data that shaped a model. The provenance of the weights. The alignment claims a lab makes about its own product. In a world of five or six compute superpowers, none of these will be verifiable by inspection, because you cannot inspect a building you are not allowed to enter.
This is the frontier I have chosen to work on โ building infrastructure that lets a model prove things about itself without revealing everything about itself. Zero-knowledge proofs over training-data provenance; attestations that survive the opacity of a proprietary dataset. The reason this matters is not ideological purity. It is that when compute concentrates, verifiability is the only remaining form of accountability. If you cannot watch the machine, you must be able to check its receipts.
Notice that this reframes the entire Inner Mongolia debate. The question the headlines asked was whether domestic silicon can match foreign silicon. The question that actually determines our future is whether any of this โ domestic or foreign, sanctioned or free โ can be made accountable to the people it affects. On that question, a deployment in a climate-controlled hall on the steppe and a deployment in a Virginia campus are equally opaque. The topology is the enemy. Not the flag.
The Pragmatism Test
I hold every technology claim to a test I borrowed from the way I audit code, and I would like to run it on this whole affair.
The test has one question: if this deployment were run by a company you did not like, in a country you did not trust, would you still call it decentralization?
If the answer is no โ if the word only applies when our friends do it โ then it was never a technical claim. It was a tribal one wearing technical clothes. I have watched this happen in my own house. I sat through the DeFi summer of 2020 and then, when the governance token crashed and took my study group's modest savings with it, I interviewed thirty people about what they actually felt. Not one of them had been told that the interest-rate curve that destroyed them was, functionally, an arbitrary invention dressed up as market equilibrium. The numbers looked objective. The curve looked discovered. It was designed, by people, with assumptions, and the assumptions were the part nobody audited.
I see the same pattern in this news. A roadmap number becomes a fact. A deployment intention becomes an installed cluster. A chip that has not been benchmarked becomes a benchmark. The objectivity is aesthetic, not structural. And the crowd that prides itself on questioning authority applies none of that skepticism when the authority agrees with it.
If you can hold the uncomfortable position โ that this deployment is simultaneously impressive, strategically rational, economically uncertain, and totally orthogonal to decentralization โ then you are finally in a position to learn something from it. If you cannot, you are just cheering.
What I Would Actually Watch
I do not traffic in predictions, and I am not going to start in a bull market. But I will tell you the few things that would move my priors, and the few that would not.
Three things would move me. First, a public, third-party measurement of tokens per second per watt on production hardware, comparing the domestic stack against the old one. That single number determines whether the economics close or need subsidy. Second, a disclosure โ from the company or the chipmaker โ that distinguishes training from inference explicitly. The silence on this point is the loudest thing in the whole story, and a company confident in its training stack would not leave it ambiguous this long. Third, evidence that the cluster was actually supplied at the stated scale, which means watching the memory supply chain, not the accelerator marketing.
Three things would not move me. A stock price. A partnership press release. Another vertical outlet relaying the same unquantified sentence with more adjectives. I have spent eighteen years in this industry and I have never once seen an adjective install a rack of silicon.
The Slow Path
I founded a small collective in 2021 because I could not stomach minting speculative profile pictures, and I hand-wrote the contract so that royalties flowed to the local artists who had actually lived the thing being sold. Fifty artifacts. No virality. I do not tell that story to be admired. I tell it because it is the only kind of work I have ever been able to defend, and because the part of me that would cheer this Inner Mongolia headline is the same part that was tempted, briefly, to mint a panda and retire.
The temptation is real. The temptation is always the same shape: something large is happening, ergo it is happening for us. I have learned to sit with the opposite feeling โ the suspicion that large things happen for themselves, and we are allowed to comment on the margin.
So here is where I land, and it is not a landing that will trend. The DeepSeek-Huawei deployment is a serious engineering event with real industrial consequences and almost no bearings on the decentralization story we tell ourselves. It belongs to the category of things that are true and not about us. The mature response is to read it for what it reveals โ that domestic silicon is close on inference, that memory is the binding constraint, that power geography is a genuine edge, that a quantitative fund's balance sheet is a strategic asset โ and then to return, without resentment, to our own hard problem: making computation that no one can inspect nonetheless accountable.
Follow the fear, not the chart. The fear here is not that America is losing. The fear is that the world is consolidating into a small number of compute islands, sovereign or corporate, and that the verifiability of what those islands do is being quietly postponed while everyone argues over their flags.
If you can build the receipts โ the attestations, the proofs, the verifiable claims that survive opacity โ then the geography of who owns the chips matters far less than it does today. That is the work I get up for at thirty-four, and it is the work I would still be doing if the chips were all American and all free.
The concrete is pouring on the steppe. Let it. Our job was never to pour concrete faster than they do. Our job is to make sure that when the dust settles, somebody somewhere can still check the work.