frontier AI labs and disruption
  • BlogBlog

Are frontier AI models overshooting the Job to Be Done?

  • FormatChris Selland
  • FormatNovember 16, 2026

A noisy debate is currently underway over whether the ballooning capital expenditures poured into frontier AI models (AI systems at the leading edge of current capability, representing the most advanced general-purpose models available at any given time) can ever generate adequate commercial returns. 

Yet much of this commentary proceeds from 30,000 feet—focusing on macro CapEx (capital expenditure) commitments and theoretical capability benchmarks without grounding the discussion in what buyers are actually procuring on the ground.

When examining emerging technology shifts, Clayton Christensen taught us to look past top-line performance metrics and ask a more fundamental question: What is the customer’s actual “Job to Be Done,” and are incumbent solutions overshooting what that Job strictly requires?

The dynamics of overshooting

In the classic framework of Disruptive Innovation, market leaders sustain their advantage by continually adding power, complexity, and features to satisfy their most demanding, high-margin customers. In doing so, their performance trajectory outpaces the baseline capability that mainstream or lower-tier users can absorb.

This dynamic creates a classic overshooting pattern:

  • The Performance Gap: The dominant product becomes too capable, complex, and expensive for everyday use.
  • The Foothold: Alternative entrants introduce simpler, narrower, or lower-cost offerings that incumbents initially dismiss as inferior.
  • The “Good Enough” Trajectory: While trailing the absolute performance frontier, these alternatives improve rapidly along dimensions overserved customers or nonconsumers value —cost, speed, and autonomy—quietly siphoning off mainstream volume.

When building a model to forecast the World Cup earlier this year, I observed this dynamic firsthand. The architecture that solved the forecasting problem was remarkably lightweight; it didn’t need a massive, multi-modal frontier model. Yet the prevailing industry narrative—largely shaped by foundation labs with a vested economic interest in massive compute clusters—remains fixated on pushing the performance ceiling.

To test whether real-world adoption reflects this overshooting pattern, I turned to empirical consumption data rather than theoretical speculation.

Measuring revealed preference: The overshoot index

For several months, I have been tracking daily AI model usage and token consumption (the total number of text units [tokens] an AI model processes during an interaction) across OpenRouter, an API routing layer that connects developers across hundreds of models and scores of independent providers. By late October 2026, the index tracks over 190 trillion weekly routed tokens.

While enterprise procurement contracts often lag by quarters, developer routing platforms serve as an agile leading indicator. They reflect revealed preference: not what lab roadmaps promise or what buyers say on surveys, but where active compute cycles and capital are actually being deployed.

Evaluating this usage across pricing tiers and architectural categories reveals three distinct structural shifts:

1. Ingestion has been thoroughly commoditized

When evaluated on input token rates, roughly 77.4% of routed network volume flows to models priced below $1.00 per million tokens. In fact, free and ultra-low-cost tiers (under $0.10 per million) command over 26% of total volume alone. Meanwhile, higher-end offerings priced at $1.00 or more capture only 13% of input traffic.

For standard ingestion, translation, context retrieval, and summarization, “good enough” has already won the high-volume battle. Mainstream builders are recognizing that paying frontier premiums simply to parse context constitutes severe over-provisioning and overspending.

2. The open-weights foothold: Disruption from alternative business models

The most telling disruption signal in the data is architectural: 72.2% of all tokens across the top 50 models flow through open-weights models (AI models whose core mathematical parameters, known as weights and biases, are publicly released and available for anyone to download).

The single most dominant offering on the network—DeepSeek’s v4.1 Flash—captures over 25% of platform volume on its own, with secondary checkpoints (backup states of an AI model that are stored on secondary, persistent, or cluster-wide storage tiers rather than fast, immediate memory) and fast-following open models (instead of spending massive amounts of capital and research time to pioneer a brand-new architecture or training breakthrough from scratch, labs [frequently led by open-weight contributors in China and Europe] rapidly reverse-engineer, train, or distill capabilities that mirror leading closed models within weeks), such as GLM-5.3 Flash and Step-5, rapidly absorbing the rest. Proprietary frontier models from OpenAI and Anthropic maintain prominent standing in the top tier, but they collectively represent a minority of total active developer traffic.

Christensen showed that Disruptive Innovations rarely enter as slightly cheaper direct clones of the incumbent product. Instead, they enter through different business models, distinct distribution channels, and modular architectures. Open-weights architectures—hosted across dozens of competing providers driving margins to cost—fit this pattern well. They turn model capability into an adaptable, low-friction commodity.

3. Output asymmetry and the dynamics of reasoning

Earlier in our tracking, an interesting divergence appeared: while input pricing collapsed, sub-dollar output volume plummeted as reasoning-intensive models entered the mix. Reasoning and multi-step inference produce vast chains of output tokens, meaning spend quickly aggregates on generation rather than ingestion.

However, recent market shifts show that pricing power in output is not an impregnable moat. As new open-weight flash architectures enter at fraction-of-a-cent output pricing, the window during which labs can command sustained output premiums is compressing. Disruption does not move in a single flat line; it advances in waves where incumbent capability leaps are rapidly met by modular commoditization from below.

What this teaches us about disruption in AI

Interpreting these market dynamics through Christensen’s framework surfaces several critical lessons for buyers and investors:

Pay attention to the lead-user signal: Why builder sophistication matters

A crucial caveat is that OpenRouter does not necessarily represent the broader enterprise software market. Its user base skews heavily toward independent software engineers, autonomous agent creators, and specialized technical teams. In short, these users tend to be considerably more sophisticated than the average corporate IT buyer.

In innovation theory, that sophistication makes them an indispensable lead-user signal rather than an anomaly to dismiss. Sophisticated developers possess the technical fluency to evaluate latency, parameter efficiency, prompt distillation, and fine-tuning. They are the first to recognize when a 70B open-weight model or an optimized “flash” checkpoint hits the “good enough” performance threshold for a specific workload—and unlike enterprise procurement organizations locked into multi-year vendor agreements, they can re-route production traffic overnight.

Historically, disruptive footholds rarely start in central enterprise procurement offices. They begin with technical practitioners building around the edges, adopting agile alternatives that traditional buyers eventually embrace once the integration paths and toolchains (a connected set of software development tools that work together to turn human-readable source code into a functional program or deployed system) mature.

Don’t confuse volume share and economic capture: Frontier labs aren’t overshooting

As developer usage shifts to fraction-of-a-cent models, the immediate temptation is to declare that frontier labs have dramatically overshot the market.

At this point in time, I don’t think that’s true.

In classic disruption theory, overshooting means you’ve built a product with capabilities nobody can absorb or justify paying for. That isn’t what is happening at the bleeding edge. If you are orchestrating complex multi-step autonomous agents, untangling messy legal contracts, or asking a model to write novel scientific code, you need every ounce of frontier reasoning power you can get. In those scenarios, high-end capability isn’t an overpriced luxury—it’s the baseline requirement to get the job done.

What we’re actually seeing isn’t an incumbent overshooting its market; it’s an explosion of choice across a whole spectrum of skills and price points. The market hasn’t rejected frontier models—it has unbundled them.

The mistake is assuming the apex model represents the whole market. Frontier labs are still capturing significant economic value because they own the hardest, highest-margin problems where buyers will pay whatever it takes for accuracy. But that high-margin apex sits on top of a massive base of everyday tasks that simply don’t require a frontier model.

Follow the Law of Conservation of Attractive Profits: Market infrastructure acquisitions are telling a story

Christensen’s Law of Conservation of Attractive Profits states that when a product’s performance becomes good enough and modularizes, the opportunity to earn attractive profits shifts to a contiguous stage of the value chain that remains proprietary and performance-defining.

If sheer frontier intelligence were the sole enduring bottleneck, strategic capital would flow exclusively to model training labs. Yet market infrastructure acquisitions tell a different story:

  • Stripe acquired OpenRouter to control the routing, execution, and billing layer connecting applications to models.
  • Nvidia committed $12.9 billion for Hugging Face to secure ownership of the central distribution, hub, and workflow layer for open weights.

Profits are steadily migrating away from raw model generation toward the orchestration, integration, and routing layers that manage how models are selected and deployed.

Bridging observed choice and true requirements

The data from developer routing platforms gives us a real-time window into builder preference. But it also points to a broader strategic reality for business leaders: the problem today isn’t that frontier labs are overshooting, but that organizations are over-provisioning.

Treating AI as a one-size-fits-all procurement decision—either buying the most expensive frontier model for every employee or settling for a generic tool that can’t do the heavy lifting—is an expensive mistake.

The real task ahead is strategic alignment: matching the specific Job to Be Done to the right tier of capability. With models proliferating across every price and performance bracket, the winners won’t be the companies that default to the biggest model available. The winners will be the organizations that build modular, flexible architectures—routing everyday tasks to fast, low-cost models and saving frontier horsepower for the complex problems that truly move the needle.

Data source: OpenRouter weekly metrics and catalog snapshots via the Overshoot Index. Analytical research and pipeline automation assisted by Claude.

Author

  • CCI Avatar
    Chris Selland

    Chris Selland is the Founder of Differential Factor LLC and a Lecturer in Entrepreneurship & Innovation at Northeastern University’s D’Amore-McKim School of Business.