Chiplets, Bottlenecks, and the 2028 AI Mirage

Huawei's Ascend 910B hits roughly 320 TFLOPS in FP16. That's within spitting distance of the A100's 312. On paper, the gap is closing faster than most Western analysts predicted. But here's the thing nobody wants to say out loud: single-card specs are the easy part. The real fight is happening in the dark corners of cluster interconnects and software stacks — and that's where the 2028 plan gets ugly. China's stated goal — training frontier AI models on domestic hardware by 2028 — is less a technical roadmap and more a declaration of war against the entire NVIDIA ecosystem. And like most declarations of war, the opening salvo looks impressive until you check the supply lines. This is the part where I remind you that I've spent years debugging smart contracts and watching promising protocols collapse under their own architecture. The pattern here is painfully familiar. Pump, dump, debug. Repeat. Let's talk about what's actually on the table. The 910C is expected to hit 70-80% of H100 performance. Cambricon's Siyuan 590 is approaching A100-level energy efficiency in training scenarios. These are real numbers, not marketing slides. But when you scale from a single card to a 10,000-card cluster, the arithmetic changes. NVIDIA's NVLink and NVSwitch, paired with InfiniBand, deliver 900GB/s+ of interconnect bandwidth. Huawei's HCCS plus RoCE? About 400-500GB/s. That's not a minor gap — that's the difference between a training run that takes three weeks and one that takes two months. The industry estimates put Chinese clusters at 70-85% linear scaling efficiency compared to NVIDIA's stack. To hit the 2028 target, they need 90%+. That's the kind of leap that doesn't happen through brute force. It requires the kind of systems engineering maturity that typically takes a decade to accumulate. And they're trying to do it in four years, under export controls that restrict access to the very advanced process nodes that make high-performance chips possible in the first place. Here's where the code-first verification instinct kicks in. I've audited enough projects to know that software ecosystems are the silent killers of ambitious hardware roadmaps. The CUDA moat isn't just about performance — it's about the decades of optimized libraries, debugging tools, and developer muscle memory that make NVIDIA the default choice. Huawei's CANN platform and MindSpore framework are improving, sure. But asking developers to migrate from CUDA to a less mature ecosystem is like asking a trader to switch from a battle-tested execution engine to a new one that might have a latency spike right when the market moves. The friction is real, and it's not going away because of a government mandate. Now, let's get contrarian for a second. Everyone's focused on the hardware — the chips, the interconnects, the HBM supply chain. But the actual bottleneck might be something far less glamorous: Model FLOPs Utilization, or MFU. Chinese clusters are estimated to hit 30-40% MFU. NVIDIA clusters? 50-60%. That means even if China builds a 100,000-card cluster that matches NVIDIA's scale, the effective usable compute is only 60-70% of what the hardware specs suggest. This isn't a chip problem. It's a systems problem — distributed training frameworks, checkpointing strategies, fault tolerance, network congestion control. It's the unglamorous plumbing of AI infrastructure, and it's where the 2028 plan looks most fragile. And let's not ignore the elephant in the room: HBM. The high-bandwidth memory that these chips depend on comes primarily from Samsung and SK Hynix. Both are subject to US export controls. Domestic HBM production, led by CXMT, is still in early stages. If the US tightens HBM restrictions — which is a real possibility given the current trajectory — the performance ceiling for Chinese AI chips gets capped regardless of how clever the chiplet packaging gets. This is the kind of supply chain vulnerability that doesn't show up in a spec sheet but absolutely determines whether the 2028 target is achievable. Gas fees higher than the yield. Typical. I keep coming back to that phrase because it captures the fundamental mismatch between ambition and execution that plagues so many projects — crypto or otherwise. The Chinese AI plan is no different. The ambition is real, the resources are substantial, and the policy support is unwavering. But the execution details — the cluster efficiency, the software ecosystem, the supply chain resilience — these are where the plan could bleed out. So what's the actual takeaway here? The 2028 plan isn't about beating NVIDIA at their own game. It's about creating a parallel ecosystem that's good enough for domestic needs and strategically independent from US control. That's a more realistic goal, and it's one that's likely to succeed in some form. The real question is whether "good enough" will be sufficient to maintain China's competitive position in AI, or whether the gap will widen as NVIDIA continues to push the frontier forward. The deeper play here is the emergence of a bifurcated global AI infrastructure — two separate compute ecosystems with their own standards, tools, and supply chains. That's not a prediction; it's already happening. The 2028 target is just the public milestone for a process that's been underway since the first export controls were imposed. The question isn't whether China will have domestic AI training capability. It's whether that capability will be a viable alternative or a strategic dead end. For the rest of the world — and for the crypto ecosystem that increasingly depends on AI infrastructure — this means paying attention to the engineering signals, not the policy headlines. Watch the MFU numbers. Watch the HBM supply chain. Watch whether developers actually migrate to the domestic ecosystem or stick with workarounds. Those are the metrics that will tell you whether 2028 is a breakthrough or a mirage. As for me, I'll be watching the cluster efficiency reports the way I watch smart contract audits: looking for the hidden assumptions that turn a promising spec into a production nightmare. Because in AI, as in crypto, the devil isn't in the details. The devil is in the systems engineering. And that's where this story will be decided. t check.