KIMI K3 vs OPENAI / ANTHROPIC KIMI is disappointingly slow. I'm going to describe why and what stocks will benefit because - guess what,- Kimi, DeepSeek, etc are all about to get A LOT better.  Here's the key difference: Kimi K3 is an extreme "mixture of experts." It's a 2.8 trillion parameter model (almost as big as Claude), but for each token it activates just 16 of its 896 experts — about 104 billion parameters. So it's "playing" on a much smaller (but hopefully just as accurate) slice of the model. Those 896 experts are spread across many GPUs. And every token can call a different set of 16 — so the GPUs are constantly passing pieces of the answer back and forth. This hammers the Interconnects.  Now assume the closed models — Claude, GPT — run dense or close to it. A dense model knows exactly where its weights live. The data flows through the same GPUs in the same order, every prompt, all planned in advance. Claude pays for that with more raw compute per token — but it barely touches the slow wires between machines. THIS IS THE BOTTLENECK OF KIMI. And for the 1000s of companies that will adopt Kimi because its cheaper. The wires between machines are the slow part. Inside a GPU, data moves at roughly 8 TB/s today. Between GPUs could be as low as 0.1 TB/s. Around 1% of the speed. It's like Kimi is riding a Ferrari and has to stop and take a bicycle every mile or so — while Claude drives one car on one highway the whole way.That's a big part of why Kimi K3 feels slow on most providers right now, while Claude feels instant. Kimi's architecture is arguably smarter per dollar of compute it's chatter runs over the slowest links in the data center. Which is why interconnects ARE THE TRADE underneath the entire open-model story: as interconnects improve, extremely sparse open-weight models are the workloads that benefit the most. Fix the plumbing and the open models' can/will surpass the closed models. The "plumbing" for 40 years was CPUs. Then GPUs. Now...Interconnects.  Who's in the race: $CRDO — the leader in copper-based electrical interconnects (active electrical cables). The incumbent technology, and the one hitting physical limits first. But copper is hitting physical limits. Too. slow.  $NVDA — just put $2 billion each into $COHR and $LITE to lock up laser-based optical interconnects. $MRVL — led the latest funding round of private Mojo Vision and is jointly developing MicroLED-based optical interconnects with them. Read that again: Marvell is betting on MicroLEDs. But no product. No MicoLED.  Avicena — private, well funded, has demoed MicroLED links but nothing scaled yet. Then there's $FABC, the highest reward play in my opinion and I'm a shareholder, which is jointly developing a high-speed, energy-efficient MicroLED interconnect (Neural I/o) with $KOPN — Big advantage: KOPN has years of MicroLED manufacturing experience, including for the military — and $FABC will own the IP with $KOPN on these interconnects. The company says it's targeting a demonstration of the platform by the end of 2026.The tell that MicroLED interconnects are real: Marvell, a $172+ billion company, just made the same bet and Jensen Huang then stated, "MRVL is the next trillion dollar company."  Kimi is painfully slow today. That's not a model problem(!) — it's a plumbing problem. And the plumbing is where the money is going.