The Chips Actually Powering This Year's AI Boom
Nvidia's Rubin, AMD's Instinct MI400, and Broadcom's custom silicon — what's actually different about each one, not what the press release says.
In this article 5 sections
Quick answer: Three chips actually moved the AI hardware conversation forward this year, and none of them are the one getting the most airtime. Nvidia's Rubin platform entered full production with real spec sheet upgrades over Blackwell, not just a rename. AMD's Instinct MI400 series is closing the performance gap without actually closing it yet. And Broadcom's custom AI silicon — the stuff nobody outside the industry can name — grew its AI revenue 221% year over year and is quietly winning more of the data center than either household name. Here's what's actually different about each one, not what the press release says.
Marketing claims about chips deserve the same scrutiny as claims about phones. Chips are harder to take apart in a useful way than a phone hinge, but the instinct applies: read the actual numbers, not the adjective the company picked for the announcement. At Protocol Wire, we cover AI, software, startups and the hardware behind them.
Nvidia Rubin: The New Baseline
Nvidia's Rubin platform moved into full production this year, and unlike some of the company's past "next-gen" announcements, the spec sheet backs up the hype. According to Nvidia's own announcement, the Rubin GPU packs 336 billion transistors and 288GB of HBM4 memory, delivering 50 sparse petaflops of FP4 inference performance against Blackwell's 20 — that's not a spec-sheet rounding error, that's a real doubling-plus on the number that actually determines inference cost at scale. The flagship NVL72 rack configuration crams 72 Rubin GPUs and 36 Vera CPUs into a single liquid-cooled enclosure pulling over 200 kilowatts. That's an absurd amount of power for one rack, and it's exactly the kind of number that should make you go back and reread the piece on what an AI data center actually costs to build and run before assuming your facility can just slot a few of these in. "New chip, same building" isn't how this generation works. The platform isn't just the GPU either — Nvidia bundles it with the new Vera CPU, NVLink 6 switching at 3.6 terabytes per second GPU-to-GPU, ConnectX-9 networking, and a newly integrated Groq 3 LPU for specific inference workloads. That's a lot of components to market as one platform, and it's also exactly how Nvidia keeps competitors chasing a moving target — AMD and the custom-silicon players aren't just competing against a GPU anymore, they're competing against an entire matched stack Nvidia controls end to end.
AMD's Instinct MI400: Closing the Gap, Not Yet Matching It
AMD formally launched the initial MI400-series SKUs, the Instinct MI455X and MI430X, built on the new CDNA 5 architecture. Coverage of AMD's own positioning puts the series at up to 40 FP4 and 20 FP8 petaflops — roughly double the prior-generation MI350, and AMD is explicitly pitching it against Rubin rather than against Nvidia's older Blackwell chips, which tells you where AMD thinks the real competition sits. Here's the part AMD's marketing won't lead with: "closing the gap" and "matching the leader" are different claims, and the MI400 series is still the second one, not the first, on raw top-end throughput. Where it actually wins is the same place AMD has quietly won for two generations now — price-per-unit-of-compute for teams that don't need the absolute frontier number and do need the budget to not blow up. AMD's own numbers back up that "value alternative" positioning: its data center revenue grew 44% year over year in the first half of 2026, a real number, just one that trails Nvidia's growth rate and comes from a business where data center made up only 57% of total revenue against Nvidia's 92%. AMD is still fighting a diversified-company discount that Nvidia, built almost entirely around data center silicon at this point, doesn't have to deal with. AMD's already talking about the MI500 for 2027, which is the surest sign yet that nobody in this industry expects the current chip to still be the answer in eighteen months.
Broadcom's Custom Silicon Quietly Wins the Data Center
While Nvidia and AMD get the headlines, Broadcom is the one actually posting the wildest growth number. The company's AI semiconductor revenue hit $16.7 billion in its fiscal third quarter, up 221% year over year, and now makes up 56% of total company revenue, according to CNBC's coverage of the earnings report. Broadcom doesn't sell you a GPU with its name on the box. It designs custom ASICs — application-specific chips built to one hyperscaler's exact workload instead of a general-purpose part sold to everybody. That's a less flashy business than "we beat Nvidia's benchmark," but it's the reason Broadcom is projecting its AI revenue to double to $115 billion in fiscal 2027 and double again to $230 billion in 2028. Broadcom's forecast is aggressive by any standard — doubling twice in two years would make its AI business alone larger than most entire semiconductor companies are today. Whether that holds depends less on Broadcom's own execution than on whether the two or three hyperscalers driving the bulk of that custom-silicon demand keep building at the same pace, which is the same concentration risk sitting underneath most of this year's AI infrastructure spending story, chip vendors included. If your mental model of the AI chip market is "Nvidia versus AMD," you're missing the company quietly building the custom silicon inside a growing share of the largest data centers on earth, with a customer's logo on the outside and Broadcom's design inside it. Another in-house accelerator, covered in how Alibaba's Zhenwu V900 changes AI chip procurement, shows the same shift from the buyer's side.
What Actually Matters When You're Buying, Not Reading a Spec Sheet
None of these three chips is "the best" in any way that matters outside a specific workload. Rubin wins on raw frontier throughput and inference cost at the absolute top end, and it demands infrastructure most facilities don't already have — see the cooling and power numbers above. The MI400 series wins if your budget can't stretch to the frontier chip and your workload doesn't need the last 15% of performance the frontier chip offers. And a custom ASIC, Broadcom's or otherwise, wins if you're a hyperscaler with one dominant workload and the volume to justify designing silicon around it specifically instead of buying general-purpose. The model running on top of any of these chips matters as much as the silicon underneath it — the benchmark gaps between this September's major model releases, covered in a roundup here on the models themselves, can move your actual cost-per-task more than a generational chip upgrade does. Buy for the workload you actually have, not the one the sales rep is pitching.
Conclusion
The chip that "wins" this AI cycle isn't going to be one part with one name on it — it's going to be whichever combination of silicon, facility, and model stack actually survives a real production workload without the invoice or the power bill blowing past what got approved. That's a less exciting sentence than "Nvidia's newest GPU is a monster," but it's the one that holds up in eighteen months, which is about how long any of these three chips get to be the newest thing before the next one shows up. It always is. Protocol Wire will keep covering this market through the numbers that matter to buyers and builders.
