$100 billion is a big number.
Broadcom CEO Hock Tan just told Wall Street that his company expects AI chip revenue to hit that figure by fiscal year 2027. Not “around” $100 billion. “Significantly in excess of” it. And as someone who spends my days building bots and wrestling with inference costs, this projection matters more than you might think.
Why Custom Silicon Matters to Your Bot Architecture
Here’s what Tan is really saying: the era of general-purpose AI chips is giving way to custom accelerators built for specific workloads. Broadcom’s AI semiconductor revenue has already more than doubled, and the company has locked in supply agreements through 2028. That kind of forward planning tells you the hyperscalers aren’t just experimenting anymore—they’re committing.
For those of us building production bots, this shift has real implications. Custom chips mean better performance per watt, lower latency, and eventually, cheaper inference. When you’re running thousands of bot interactions per hour, those margins add up fast.
What the Numbers Actually Mean
Broadcom’s fiscal Q1 2026 results showed AI chip revenue surging 106% to $8.4 billion, with total revenue climbing 29% to $19.31 billion. Do the math on that trajectory, and Tan’s $100 billion projection starts to look less like hype and more like extrapolation.
But here’s the part that matters for bot builders: this isn’t about training models. Broadcom’s strength is in custom accelerators—the chips that run inference at scale. That’s the bottleneck most of us actually face. Training happens once. Inference happens millions of times.
The Supply Chain Signal
Broadcom securing supply through 2028 is the quiet part that deserves attention. Chip fabrication doesn’t scale overnight. When a company locks in manufacturing capacity three years out, they’re betting on sustained demand, not a temporary spike.
For developers, this means the infrastructure we’re building on today will likely get better and cheaper, not scarcer and more expensive. That’s the opposite of what happened during the crypto mining boom, when GPU prices went through the roof and availability tanked.
What This Means for Your Next Bot Project
If you’re architecting a bot system right now, you’re probably making decisions based on current cloud pricing and available hardware. But Broadcom’s numbers suggest that space is about to shift significantly. Here’s what I’m thinking about:
- Inference costs should trend downward as custom silicon reaches scale
- Latency improvements will make real-time bot interactions more viable
- Edge deployment becomes more practical with specialized chips
- Multi-modal bots (text, vision, audio) get cheaper to run
The Bigger Picture for AI Infrastructure
Tan’s projection isn’t just about Broadcom. It’s a signal about where the entire AI hardware market is headed. When one major player sees this kind of growth, it means the hyperscalers—your AWS, Azure, and Google Cloud—are placing massive orders for custom chips.
That investment flows downstream to us. Better chips mean better APIs. Better APIs mean we can build more capable bots without burning through our budgets on compute costs.
A Reality Check
Should you redesign your entire bot architecture based on one CEO’s revenue projection? No. But should you pay attention when a major chip supplier says they’ve secured supply for the next three years and expects revenue to grow more than tenfold? Absolutely.
The bot builders who win over the next few years will be the ones who understand not just the software layer, but the hardware economics underneath. Broadcom’s numbers suggest those economics are about to get a lot more favorable.
For now, I’m keeping my architecture flexible and my eye on inference pricing. When those custom accelerators hit scale production, the bots we can build—and afford to run—are going to look very different from what’s possible today.
đź•’ Published: