The corporate PR machine works on a simple loop. A partner whispers about a slowdown, a hyperscaler adjusts its capex cadence, and the hardware ecosystem panics. Then comes the executive shield. Broadcom chief Hock Tan stepped up to the podium and swore on silicon that their artificial intelligence revenue targets remain completely untouched despite Anthropic allegedly easing off the accelerator.
The market exhaled. Wall Street cheered. For an alternative view, read: this related article.
Everyone missed the point.
Focusing on whether Broadcom hits an
Broadcom Is Missing The Real AI Meltdown And It Will Cost Them Billions Related reporting on the subject has been published by Ars Technica.
technology, business
Hock Tan is playing a very expensive game of chicken with math, and Wall Street is cheering him on because people love a clean growth story. When the Broadcom chief executive casually dismissed Anthropic's reported push to pump the brakes on scaling, insisting that custom silicon orders and data center expansion targets remain entirely untouched, the mainstream tech press nodded along. The lazy consensus is simple: hardware demand is decoupled from software efficiency, because hyperscalers have infinite capital and an unquenchable thirst for compute.
That theory is dead wrong.
I have watched enterprise technology budgets bleed out in real-time over the last twenty years. Whenever software efficiency gains outpace hardware deployment speeds, the hardware bottleneck vanishes overnight. Anthropic and OpenAI are not slowing down because they suddenly found religion or want to save electricity. They are hitting the structural wall of diminishing returns on raw parameter scaling. When training efficiency improves by a factor of ten, you do not need ten times the silicon. You need one-tenth of it. Broadcom is betting its custom ASIC pipeline on a continuation of brute-force scaling that is already dying under its own weight.
Let us look at the mechanics of what is actually happening inside the modern inference cluster. Broadcom makes its billions tying hyperscalers to proprietary networking fabrics and custom accelerators. They provide the plumbing for the artificial intelligence boom. Their business model relies on the assumption that every incremental drop in model loss requires an exponential increase in watts and silicon acreage.
Except algorithmic breakthroughs are starting to eat hardware margins alive. State-space models, mixture-of-experts architectures with aggressive routing, and post-training compression techniques are squeezing the computational fat out of these models faster than TSMC can etch a three-nanometer wafer. When an AI lab figures out how to run a model with seventy percent fewer active parameters while maintaining performance, the forward order books for custom silicon do not just plateau. They contract.
Tan knows this, of course. Executives of his caliber do not miss structural shifts; they manage expectations while positioning their balance sheets for the pivot. But the analyst community is high on its own supply, treating Broadcom's backlog like an annuity bond backed by the federal reserve. It is not. It is a cyclical hardware bet masquerading as a structural secular trend.
Consider the reality of data center power constraints. You cannot plug another fifty megawatts of silicon into a Virginia substation just because Jensen Huang or Hock Tan built a prettier presentation. The grid is tapped out. When frontier labs talk about a slowdown, they are often translating power grid realities and thermal dissipation walls into corporate diplomacy. If you cannot power the cluster, you stop buying the chips. It really is that simple. Broadcom’s backlog looks secure on paper because purchase orders are sticky until the moment penalties for cancellation are cheaper than taking delivery of hardware you cannot plug in.
The mainstream narrative treats custom accelerators like Google's TPU or Meta's MTIA as permanent replacements for merchant silicon. The assumption is that hyperscalers will keep spinning their own chips forever to avoid paying Nvidia and Broadcom a toll. That misses the underlying economics of chip design. Designing custom silicon only makes sense at hyper-scale when your workload is static and your algorithm is locked in stone. But the architecture of artificial intelligence changes every six months. If you tape out a custom chip for a transformer architecture that gets leapfrogged by a radically different sequence model before the silicon even returns from the foundry, you just burned three hundred million dollars on a paperweight.
This is the vulnerability nobody wants to price into Broadcom's valuation. Their custom ASIC business is tethered to massive upfront engineering investments by companies that are starting to realize their software stacks are shifting too fast for hardware lock-in. When Anthropic or OpenAI alter their training paradigms to prioritize sparse compute and dynamic routing, the utility of fixed-function custom silicon drops off a cliff.
People love to ask whether enterprise artificial intelligence spending is a bubble. That is the wrong question entirely. The spending is real, but the composition of that spending is about to mutate violently. Enterprises are already figuring out that they do not need frontier models running on clusters of ten thousand accelerators to automate their workflows. They need smaller, highly quantized, domain-specific models running on edge hardware or lean cloud instances.
When the market shifts from training massive general-purpose monoliths to deploying hyper-efficient specialized inference engines, the massive custom silicon gravy train slows down. Broadcom will still sell networking gear, because data centers always need switches. But the high-margin, hyper-growth custom accelerator bonanza that justifies their current multiple relies on a continuation of the brute-force arms race.
And that race just hit a speed bump called physics, followed closely by economics.
Stop looking at the backlog. Start looking at the power bills, the algorithmic efficiency gains, and the rate at which custom silicon designs get abandoned before tape-out. Hock Tan can say whatever he wants about revenue targets, but he is selling shovels in a gold rush where the miners are suddenly realizing they are digging in dry dirt.
