The MHz Wars of AI

There’s a pattern that repeats itself in the history of computing, and we might be living through its AI equivalent right now.

In the early CPU era, the benchmark that mattered was clock speed. Intel vs. AMD fought on megahertz, then gigahertz. Marketing budgets were built around single numbers. Consumers learned to equate “faster clock = better chip.” It was clean, legible, and — for a while — true.

Then it stopped being true. The gains from raw frequency hit physical and thermodynamic walls. The industry pivoted almost overnight: instead of faster cores, build more cores. Parallelism became the new religion.

We’re watching the same movie play out in AI.

For the past few years, every major lab has raced on a vertical axis — reasoning benchmarks, MMLU scores, context windows. GPT-5.5 vs. Claude Fable 5 vs. Gemini vs. whatever dropped last Tuesday. The implicit promise is the same one Intel made in 1999: a bigger number on the box means a better product.

But cracks are forming.

The Multi-Core Moment

OpenRouter’s Fusion — released June 12, 2026 — is the industry quietly admitting what parallel computing admitted two decades ago: the single-core ceiling is real.

Fusion works in three steps. First, a prompt is dispatched to a panel of models simultaneously, each equipped with web search and tool access. A judge model then compares their outputs, mapping consensus and contradictions. Finally, a frontier synthesizer — Claude Opus 4.8 — writes one unified answer. Roughly 75% of the performance gain comes from that synthesis step alone; only 25% comes from model diversity.

The results, benchmarked on DRACO — a deep research evaluation by Perplexity covering 100 tasks across ten domains including law, medicine, and finance, with negative scoring for wrong answers — tell a striking story:

Configuration	DRACO Score
Fable 5 + GPT-5.5, fused by Opus 4.8	69.0%
Claude Fable 5 solo	65.3% (93/100 tasks scored)
Budget panel fused (Gemini 3 Flash + Kimi K2.6 + DeepSeek V4 Pro)	64.7%
Solo GPT-5.5	60.0%
Solo Claude Opus 4.8	58.8%

A committee of mid-tier models — none individually above 60% — collectively outpaces the reigning frontier standards from both OpenAI and Anthropic, at a fraction of the cost. When a panel can do that, you have to ask what “frontier” even means anymore.

The Independence Problem

The theoretical backbone here is Condorcet’s Jury Theorem from 1785: if each voter is correct more than half the time and votes independently, collective accuracy approaches certainty as the group grows. Different models satisfy the competence condition naturally — they each clear 50% individually. But the independence condition is where things get complicated.

LLMs trained on overlapping internet scrapes share the same blind spots, the same hallucinations, the same cultural biases baked into Common Crawl. A 2026 study found that a 9-judge LLM panel effectively provided only 2.0–2.5 independent votes — three-quarters of its nominal independence wiped out by correlated errors. When errors correlate, Condorcet’s guarantee collapses. You’re not getting nine juries; you’re getting the same juror, nine times.[1]

This is precisely why Fusion’s architecture is more subtle than “more models = better.” The budget panel’s diversity — different training pipelines (Google, Moonshot, DeepSeek), different architectures, different post-training regimes — is engineered to maximize genuine independence, not just headcount. And crucially, the synthesizer doesn’t vote. It reads all outputs as a meta-reasoner, identifying which reasoning chains are structurally strongest rather than which answer appears most frequently. That shifts the mechanism from majority voting to editorial judgment — and sidesteps the correlation problem entirely.

The Latency Wall

Before declaring the multi-core moment a fait accompli, there is an architectural constraint the performance table doesn’t show: time.

Running a panel means fan-out latency — dispatching to three or four models in parallel and waiting for the slowest one to finish before synthesis can begin. That worst-case tail latency compounds with a full Opus 4.8 synthesis pass on top. For a deep research query, this is tolerable. For a customer service chatbot, a coding copilot with an impatient developer, or anything with a sub-second UX requirement, it is prohibitive. This isn’t a footnote about edge cases — it’s the primary architectural barrier separating compound models from replacing single models wholesale. The multi-core CPU was transformative, but real-time control systems still run on single-threaded loops. Compound AI will likely follow the same partition: asynchronous, high-stakes tasks absorb the latency budget; synchronous, low-stakes interactions stay single-model.

The Economic Forcing Function

Here’s the dimension that makes the architectural shift irreversible: economics.

When a budget panel matches a frontier model at half the cost, the obvious reading is “same quality, cheaper.” But the more consequential implication runs through Jevons’ Paradox — the 19th-century economic observation that when a resource becomes more efficient, total consumption of that resource tends to rise, not fall, because cheaper access unlocks entirely new demand. William Stanley Jevons noticed it first with coal and steam engines: more efficient boilers didn’t reduce coal consumption, they made coal-powered industry economically viable at scales that were previously unthinkable, and total coal demand exploded.[2]

The same logic applies here. When intelligence costs half as much, organizations don’t run the same queries and pocket the savings. They run twice as many queries, spin up compound pipelines they couldn’t justify before, and route the freed compute budget into massive parallel fan-outs that were previously too expensive to contemplate. Cheaper inference doesn’t decrease AI consumption — it multiplies it. Orchestration becomes not an alternative to frontier models but an architectural mandate: the natural destination of a world where intelligence is nearly free and the bottleneck shifts entirely to coordination, synthesis, and routing quality.[3]

So What Comes After the Core Wars?

Here’s where the CPU analogy gets interesting — and a little humbling for LLM labs.

After multi-core became table stakes, the CPU stopped being the hero. It became infrastructure. Background. The component your OS scheduler worries about so you don’t have to. The GPU took center stage — not because it was smarter than the CPU, but because it was shaped differently for what the world actually needed to do next: massively parallel, specialized computation at a scale general-purpose cores were never designed for.

The question worth sitting with: what is the GPU of the AI stack?

Two Candidates

One candidate is multimodal capability — not the halfhearted “here’s an image attachment” kind that exists today, but something architecturally different: a model that tracks the physical trajectory of objects across a video stream, reasons about the spatial geometry of a room from sensor data, or generates synchronized audio-visual output from environmental context rather than from text prompts. These are tasks that require continuous, cross-modal state — something a token-prediction engine is structurally ill-suited for, the same way a CPU’s serial pipeline was ill-suited for rendering millions of pixels in parallel.

The other candidate is what Yann LeCun has been advocating for years — world models. His argument is that LLMs are fundamentally autocomplete engines, and autocomplete, however sophisticated, hits a hard ceiling when it comes to agency and causal reasoning. A world model that maintains persistent internal representations of how things work — physical dynamics, social cause and effect, long-horizon planning — would be architecturally different from a transformer, the way a GPU is architecturally different from a CPU. It wouldn’t replace LLMs; it would sit above them, using them as a language interface while handling the reasoning the transformer was never built to do.

There is a third candidate worth naming: the orchestration layer itself. If frontier output now belongs to systems rather than individual models, the moat shifts from who trained the biggest model to who routes, synthesizes, and manages the best collection. OpenRouter’s Fusion scored 69.0% — higher than any single model at any price point. No individual lab can sell you that. The aggregator who owns the routing and synthesis layer may end up being the GPU of this era: not the smartest component in the stack, but the one everything else runs through.

What This Means Now

Predicting which paradigm wins is roughly as reliable as picking the next dominant frontend framework. The Fusion results carry their own caveats: DRACO measures written deep research, a task where synthesis naturally shines. Whether compound architectures extend to agentic work, real-time interaction, or creative tasks with no correct answer remains genuinely open.

But the shape of the transition is familiar enough to be instructive.

The companies betting everything on raw model capability are in the position of a CPU manufacturer in 2005 — not wrong, but about to be surprised by where the leverage moves. The interesting bets are on whoever figures out what task the world actually needs done that current LLMs are poorly shaped to handle, and builds the specialized architecture to handle it.

The MHz wars were exciting while they lasted. But the real computing revolution came from whoever stopped trying to make one thing faster, and started asking: what kind of thing do we actually need?