Meta’s plan to break its dependence on Nvidia just hit a wall. The company bought chip startup Rivos to speed up its in-house AI silicon, but according to The Information, the effort isn’t delivering the acceleration Meta hoped for. That’s a meaningful stumble for a company spending tens of billions a year on AI infrastructure, and it tells you something about how hard this race actually is.
Here’s why it matters. Every major AI player wants off the Nvidia treadmill. Nvidia GPUs are expensive, supply is tight, and when one vendor controls your most critical input, your margins and your roadmap belong partly to them. The Information reports that Meta saw Rivos as a shortcut to its own competitive chips. So far, that shortcut looks more like a detour.
📉 What’s actually going wrong
Buying a chip team and shipping a working data center accelerator are two very different things. Custom silicon is one of the hardest problems in tech. It’s not just the chip. It’s the software stack, the compiler, the networking, and the years of tuning that make Nvidia’s CUDA ecosystem so sticky.
Meta has been down this road before with MTIA, its Meta Training and Inference Accelerator. Progress there has been slow and uneven. Folding in Rivos was supposed to inject fresh talent and a more mature design. Integration friction, competing internal roadmaps, and the sheer difficulty of matching Nvidia’s performance per watt appear to be slowing things down instead.
🏗️ The bigger pattern
Meta isn’t alone, and that’s the real story. The hyperscalers have split into two camps:
- Google has spent roughly a decade on its TPUs and now runs much of its AI on them. That’s the proof that in-house silicon can work, but it took years and enormous patience.
- Amazon built Trainium and Inferentia and is pushing customers toward them to cut Nvidia reliance.
- Microsoft has its Maia chips, also early and unproven at scale.
- Meta is arguably furthest behind among the giants, despite spending like a leader.
What stands out here is the gap between capital and capability. Money buys startups and fabs. It does not buy the institutional muscle Google built over ten years. Meta is learning that lesson in public.
💡 Why this is significant now
Meta has committed to a staggering buildout, with capital expenditures heading toward the $100 billion range as it chases superintelligence. The more of that budget that flows to Nvidia, the thinner the strategic payoff. Custom chips were meant to bend the cost curve over time. If Rivos can’t accelerate that timeline, Meta stays locked into buying Nvidia at scale, with all the cost and supply risk that brings.
There’s also a talent angle. Meta has been paying enormous sums to staff its superintelligence push, and reports of internal turmoil have been piling up. A chip program that stalls after an acquisition adds to the sense that throwing money at the problem isn’t translating into execution.
🧭 Practical takeaways
If you’re building or buying AI infrastructure, a few lessons land:
- Don’t assume an acquisition buys speed. Chip and deep-tech teams need integration time measured in years, not quarters. Budget for that.
- The software moat is the real moat. Nvidia’s dominance is as much CUDA as it is silicon. Anyone betting on alternative chips should weigh the porting and tuning cost, not just the sticker price.
- Watch Google as the benchmark. TPU is the working model for vertical integration. The question for Meta and others is whether they can compress a ten-year journey into a few.
What comes next is the test. Meta can keep pouring money in, poach more chip talent, or quietly lean harder on Nvidia while it sorts out the in-house program. None of those paths are clean. For now, the takeaway is sober: in AI silicon, capital is the easy part. Execution is where the giants separate, and Meta has work to do. You can find the full reporting at The Information.