I have a habit that borders on compulsion: I read claims the way auditors read ledgers. In 2017, during my final years in Copenhagen, I spent six months dissecting the whitepapers of forty-one initial coin offerings. Forty-one. I kept a spreadsheet, a meticulous artifact of promises and omissions. For each project I logged the stated value proposition, the consensus mechanism, the token distribution, and — perhaps most tellingly — the gaps: what they chose not to say. I timed my reads. Fourteen minutes per document, on average. The conclusion was always the same. The noise of grand claims travels faster than the signal of actual parameters, and in that velocity gap, entire markets were built and dismantled.
That old discipline resurfaced with violence when I encountered the recent coverage of Alibaba's rumored new AI accelerator — breathlessly described in industry headlines as "China's most powerful AI chip." The coverage has a peculiar anatomy: a headline with a pulse, and no body. No process node. No yield rate. No confirmed foundry. No HBM generation. No floating-point throughput. No memory bandwidth. No launch timeline. No customer. No price. A claim stripped of every measurement that would make it meaningful.
"Strongest," "most powerful," "breakthrough" — these are not engineering terms. They are liturgical ones. And I have learned, through years of watching both crypto markets and supply chains convulse, that liturgical language deserves special skepticism precisely because it feels so satisfying.
We built the temple, but forgot who the god is.
This is not merely a semiconductor story. It is a trust story — written in silicon, export controls, and memory embargoes. And I will treat it as such.
Context: The Ledger of Production
To understand what Alibaba's chip claim actually means, we have to trace its lineage — both the engineering lineage and the regulatory one.
The engineering lineage begins in earnest in 2019, when Alibaba's semiconductor arm, T-Head (Pingtouge), unveiled the Hanguang 800, a 12nm inference NPU designed for the company's internal cloud workloads. It was a competent chip, optimized for the kind of inference traffic that Alibaba Cloud actually serves — but it was emphatically not a frontier accelerator. In 2021, the company released the Yitian 710, a 5nm ARM-based server CPU fabricated, crucially, at TSMC. That foundry choice matters: it tells us that when supply chains were open, Alibaba used the best manufacturing available on the market. The company is not an autarkic semiconductor effort. It is a fabless design house that participated in globalization while it lasted.
Now the new product, still unnamed in credible disclosures, arrives at a moment when globalization no longer exists. Since October 2022, the U.S. Bureau of Industry and Security has escalated export controls on advanced computing in waves. The first wave targeted GPU-class accelerators above specific interconnect and compute thresholds. The second, in 2023, closed loopholes around memory bandwidth and board-level designs. The third, in 2024, explicitly targeted high-bandwidth memory itself. Meanwhile, the Dutch government has blocked ASML from delivering EUV lithography systems to mainland China and restricted the most advanced DUV immersion systems. Japanese controls constrain critical materials — photoresists, etching chemicals, cleaning equipment.
The result is a manufacturing environment that no amount of design brilliance can fully circumvent. To produce something that qualifies as "advanced" in 2025, a Chinese chip designer must rely on SMIC's N+2 process — a DUV-multipatterning approximation of 7nm-class geometry, without EUV. DUV multiple patterning is technically feasible but economically punishing: more mask layers, more defect risk, more process steps, and a yield curve that matures painfully slowly. If the headline "China's most powerful AI chip" is to be taken seriously, this is the substrate on which it must be built.
There is a deeper point buried in this context that I want to bring forward: the supply chain is a ledger. Everywhere a product acquires an input — a lithography step, a memory die, a packaging interposer — an entry is written. When that ledger is policed by export controls, what is being audited is not just hardware, but the very ambition of a country to compute at the frontier. The story of Alibaba's chip is, at its core, a story about which entries can and cannot be made in that ledger.
Core: The Architecture of Constraint
The yield curve is the real governance mechanism
Let me start with the most underappreciated constraint: yield rates.
In my corner of the world, we have a phrase: "code is law." I have spent a decade probing its limits. In my 2018 essay "Code as Constitution," I argued that smart contracts, like constitutions, encode values that their authors rarely acknowledge. The same principle holds, with eerie precision, for semiconductor fabrication. In crypto, the equivalent would be: "the whitepaper is not the protocol." In silicon, the version that matters is starker — tape-out is not truth.
A chip can exist as a physical artifact in a laboratory, performing beautifully under controlled conditions. None of that matters until it yields — at scale, at acceptable cost, across millions of instances. And here, the data paints a grim picture for the "strongest" narrative.
SMIC's N+2 process — again, the most likely candidate for any advanced Alibaba accelerator — is widely estimated by independent industry analysts to yield around 50% or below for complex logic dies. TSMC's 7nm node, by contrast, achieved mature yields above 90% years ago. A 50% yield does not merely double the cost per good die; it also constrains capacity allocation. Every wafer that produces a defective die is a wafer that produced nothing usable. At 50% yield on a multi-die design, the probability of a fully functional final product compounds into single-digit territory unless redundancy and repair logic are carefully engineered.
This is also why sharing matters. SMIC's advanced capacity is rationed across multiple domestic customers: Huawei's Ascend series, Cambricon, and Alibaba, among others. The scarce resource in Chinese AI is not design talent. It is a finite set of exposure fields inside a lithography tool that must be managed like a contested blockchain — except the blocks are wafers and the miners are DUV scanners running double and quadruple patterning.
I learned this lesson the hard way in 2020. During DeFi Summer, I interned with a small Copenhagen-based DAO working on lending protocols. My task was to investigate the real-world implications of algorithmic stablecoins. I interviewed twelve users who had lost savings to oracle failures — some who had lost their entire emergency funds because a price feed reported one number while the market traded another. The pattern I took from those interviews is this: systems collapse not at their strongest point but at their tightest coupling. A yield rate is an oracle for your fabrication process. When it reports bad data, everything downstream compounds.
HBM: the nervous system is embargoed
And this brings me to the most fatal constraint of all, one that almost all commentary underweights: HBM. High-bandwidth memory.
Here is the uncomfortable geometric fact of modern AI accelerators. Compute density is only half the equation. An AI training workload — let us say, a dense Transformer run — consumes data faster than any conventional memory subsystem can feed it. The solution, adopted universally by leading accelerators, is HBM: densely stacked DRAM, mounted vertically on a silicon interposer, delivering absurd bandwidth at close physical proximity to the compute die. NVIDIA's current frontier accelerators are unthinkable without HBM3E.
The global supply of HBM is controlled by three companies: SK hynix, Samsung, and Micron. All three have effectively been restricted from selling their advanced HBM generations into the Chinese market. The U.S. rule changes of 2024 explicitly targeted these memory products. China's domestic alternative, CXMT, is making genuine progress but remains a generation or two behind on density, bandwidth, and stack height. When you multiply the performance gap by the yield gap, the effective memory disadvantage is enormous.
Let me be direct: a chip equipped with HBM2-class memory, or a constrained allocation of HBM3, does not compete with an HBM3E-equipped frontier accelerator — regardless of the quality of the compute architecture.
You can assemble a world-class design team. You can match NVIDIA's architectural innovation on paper. But the memory subsystem — the nervous system of the machine — is embargoed. This is why, in my analysis, the "strongest" claim encounters a hard ceiling that no amount of Chinese engineering ingenuity can raise quickly. Time, capital, and export control policy all stand at the door.
I recall, during my work in 2024 bridging AI and blockchain communities, co-organizing three workshops with fifty participants each, demonstrating how zero-knowledge proofs could protect AI training data privacy. The recurring sigh across all three workshops was the same: the bottleneck was never the algorithm, never the cryptographic scheme. It was the availability of the substrate. ZK proofs, like AI training, are ravenous consumers of compute and memory. When the machines are constrained, the ideas stay on the drawing board.
Advanced packaging: the visible invisible
There is yet another gate, less discussed in Western commentary but equally limiting: advanced packaging. Frontier AI chips are no longer simple monolithic dies. A modern accelerator is a small city of chiplets, memory stacks, and interconnect fabric, assembled on an interposer using techniques like CoWoS. TSMC's domination of this domain is so complete that it has become a bottleneck even for NVIDIA — which frequently allocates its CoWoS supplies among customers.
In China, packaging companies — notably JCET and Tongfu Microelectronics — have developed what one might charitably call CoWoS-like capabilities. Proximity exists; equivalence does not. Capacity is limited, yields are unproven at scale, and the most advanced 3D stacking processes remain out of reach.
So we arrive at a triple bind: the design may be sound, but the manufacturing node is constrained; the memory that the design demands is embargoed; and the packaging that binds everything into a coherent system is scarce. A chip is only as "strong" as its weakest binding constraint. In a supply-chain sense, Alibaba's chip is simultaneously a triumph of design and a hostage of infrastructure.
Code is law, until the law breaks the code. Here, the law — export control rules written in Washington, The Hague, and Tokyo — has broken something more elemental than a software provision. It has fractured the global supply chain itself. We are no longer looking at a single global ledger of production. We are looking at two parallel ledgers, with different entries, different truth conditions, and different paces of advancement.
The money trail: follow the capital
Let me turn now to the economic structure, because the capital allocation tells us more than any press release.
In February 2025, Alibaba announced a massive capital expenditure program: roughly RMB 380 billion — about $53 billion — over three years, aimed at AI infrastructure and cloud computing. The number, in isolation, looks like a semiconductor war chest. But in my experience — and I say this as someone who has spent years studying how capital flows distort mission statements — allocations matter more than magnitudes.
The overwhelming share of this capital is destined for data centers, networking, and general compute infrastructure. Not for self-built wafer fabrication. Alibaba has not announced a fab. It has not announced a captive HBM production line. The company is pursuing, as far as the disclosures show, a "design plus procurement" model.
This is the Google TPU model, and understanding that fact rewrites the entire strategic assessment. Google designs its own TPUs primarily to optimize cost-efficiency and control over its internal infrastructure — not to challenge NVIDIA in the merchant market. Alibaba, by analogy, designs its own accelerators to reduce procurement dependence, to sharpen negotiation leverage with external suppliers, and to secure a supply floor in a world of export controls. The commercial objective is not conquest. It is resilience.
The strategy is rational and, in many ways, admirable. But it is not the strategy the headlines imply. "Challenging U.S. dominance" is a narrative. Reducing supply-chain risk is a strategy. The two are frequently confused, in part because the former makes better copy, and in part because the companies involved have little incentive to correct the confusion.
I published a monthly newsletter throughout the 2022 bear market called "Quiet Crypto" — deliberately positioned against the hype of a collapsing market. The discipline I cultivated was simple: follow the money, not the narrative. The money in this story says Alibaba's chip program is a defensive instrument, not a world-domination engine.
The uncomfortable comparator: Huawei
Now we arrive at the detail that the coverage tends to omit, and the omission tells its own story.
In the domestic Chinese accelerator market, Alibaba is not the leader. Huawei's Ascend line occupies that position — deployed more broadly across Chinese clouds, driven by more extensive state-backed procurement channels, and strengthened by years of iteration under sanction pressure. The gap is not trivial. Ascend has achieved a level of ecosystem maturity, software tooling, and domestic adoption that no other Chinese accelerator has matched.
The "strongest chip" narrative, as constructed in the reporting, appears to exclude Huawei from the comparison set. That is a selection-bias problem with real consequences. If the phrase "China's most powerful AI chip" silently assumes Alibaba is the benchmark, the headline is not informative; it is flattering.
In my analysis of the 2022 DeFi failures, I encountered the same pattern repeatedly: somebody comparing a cramped protocol to the weakest alternatives rather than the strongest baseline. The output was reassuring — and wrong. The same logic corrupts the semiconductor narrative. The correct question is not "Is Alibaba's chip better than what Alibaba previously made?" It is "Is Alibaba's chip better than Huawei's Ascend, and does it approach the frontier that NVIDIA defines?" Those are very different questions, and only one of them has a plausible affirmative answer.
The CUDA cathedral
Finally, I have to speak about the software ecosystem, because this is where "strongest" meets its most existential challenge.
NVIDIA's CUDA is not a toolchain. It is a cathedral — a towering, decades-old accumulation of libraries, kernels, profilers, debugging tools, and trained human intuition. Every major AI framework — PyTorch, TensorFlow, JAX — has been optimized, hardened, and battle-tested on CUDA for years. Migrating a production AI workload from CUDA to any alternative is not an engineering task; it is a theological conversion. The cost is measured not merely in engineer hours but in reproducibility, numerical parity, and the quiet risk that your models will behave differently on a different substrate and nobody will know exactly why until the damage compounds.
The domestic Chinese ecosystem is improving. Compiler efforts, compatibility layers, and framework adaptations are real and accelerating. But a chip is only as strong as the software that allows real teams to run real workloads on it. Faith in the protocol is not faith in the people. Similarly, faith in a hardware specification is not faith that the software stack exists, that the debugging utilities are mature, or that a graduate student in Shanghai can get the same performance with the same effort as a graduate student in Palo Alto.
In my vision of ethical AI — the work that has animated me since the "Trusted AI on Chain" whitepaper I co-authored in 2024 — the software layer is the trust layer. Hardware is the physical substrate, but the trust contract is written in software. Until the software ecosystem around a chip is as credible as the silicon itself, the chip remains a promise awaiting execution.
Contrarian: The Pragmatism Test
There is a contrarian reading of all this that most coverage — both Western and Chinese — prefers to avoid.
What if Alibaba isn't trying to beat NVIDIA at all?
Think about it. A vertically integrated cloud provider that designs a meaningful fraction of its own chips gains leverage in procurement conversations. It can walk into a negotiation for external accelerators with a credible internal alternative. Not an equivalent one — but a credible one. The threat alone changes prices and allocation terms. This is precisely how sophisticated enterprises manage their supply chains, and how sovereign states manage their alliances. The chip is a negotiating instrument, not a declaration of war.
This reading also explains the vagueness of the public claims. A chip that is a strategic hedge does not need to be benchmarked against NVIDIA; it needs to be visible enough to be credible, and measured enough to be deniable. It is a political token as much as a technical object, minted for domestic audiences, procurement officers, and international negotiators alike.
And yet, the deeper contrarian point is more uncomfortable, and I reach it now from my own values rather than from the data — because the data, as I have established, is too thin to settle anything on its own.
Even if Alibaba's chip succeeds on every technical metric. Even if it achieves parity with the frontier. Even if it rolls out across Alibaba Cloud at massive scale, drives down inference costs, and powers the next generation of Chinese models. Even then, the architecture of control remains centralized. One company would own the silicon, the cloud, the model weights, the user data, and the policy decisions about what gets deployed.
That is not decentralization. It is a change of who sits at the top of the stack.
I built my career on the belief that blockchain's true power lies in encoding democratic values into immutable logic. The Chinese sovereign narrative and the American corporate narrative both converge on an uncomfortable endpoint: concentrated control over the most consequential computation infrastructure in history. For the crypto community to applaud either as a victory would be to mistake a change of temple for a change of god.
We traded soul for speed, and called it progress.
The blockchain community did not build a new internet so that we could shuffle which fences the pilgrimage passes through. If the values I have spent a decade articulating — transparency, user sovereignty, verifiability — have any meaning, they apply equally to chips. The supply chain should be auditable. The deployment should be accountable. The computation should be verifiable — not by a company's public relations department, but by mechanisms robust enough to survive state pressure and market incentives.
Takeaway: The Entries That Matter
So what do we actually watch from here? I would propose three data points, visible within a twelve-month window, that will tell us more than any press release ever could.
First: does this chip appear in Alibaba Cloud's production catalog — not as a demonstration, not as a benchmark sample, but as a deployable unit that real customers can provision, with documented performance under real workloads? Absorption into production is the only honest test of a chip's claim to power.
Second: does CXMT's HBM progress meaningfully close the bandwidth gap, and does the domestic packaging ecosystem — JCET, Tongfu, and others — scale its CoWoS-like capacity to match? These two signals will determine whether any Chinese accelerator can escape the gravity of its own supply chain.
Third: what does the next iteration of U.S. export control policy restrict? The regulatory calendar is the true ledger of this era. Every new rule writes entries that no amount of announcement rhetoric can overwrite.
The industry will continue its liturgy of "strongest" and "most advanced." I would humbly suggest a different measure: not performance parity, but autonomy parity. Not TOPS, but trust. A chip is a promise. A supply chain is the audit trail of that promise. And a ledger is only as sacred as the claims it can verify.
The ledger remembers, but the heart forgets.
Let us remember.