The Squeeze: How AI Providers Are Cornering Enterprise Teams — and the Open Ecosystem That Breaks Free
There is a slow squeeze happening in enterprise AI, and most teams are not paying attention to it. Model providers are systematically tightening the conditions under which you can use their technology — shorter sessions, escalating costs for capabilities that used to be included, proprietary tool ecosystems that only work with their models, and behavioral data collection baked into the tooling you depend on every day. Meanwhile, the open-source model ecosystem is delivering frontier-class performance that runs on hardware you already own. This is the moment to reassess who you are building on.
the tokenization of developer budgets
For years, flat-rate AI subscriptions made budgeting simple. That era is ending. GitHub Copilot — the most widely deployed AI coding tool in the world — switched from flat-rate to per-token billing on June 1, 2026. Developers immediately reported costs jumping from approximately $29 per month to $750 per month, and from $50 per month to $3,000 per month — not for more usage, but for the same usage they had before under a different pricing structure.
TechCrunch documented the broader pattern in a June 2026 analysis they called "the Tokenpocalypse": the deeply subsidized pricing that drove AI adoption was never sustainable. Enterprise AI providers spent years underpricing access to establish lock-in, and now that teams have built workflows, CI/CD pipelines, and internal tools around these APIs, the pricing floor is rising. Uber reportedly burned through its entire AI budget allocation in four months as token costs scaled. The subsidy era is ending, and the teams who built on a single provider are discovering that switching costs are high precisely because they were designed to be.
proprietary toolchains and behavioral data collection
Pricing is the visible part of the problem. The subtler part is what happens to your behavioral data when you use vendor-provided development tooling.
Anthropic's Claude Code collects session behavioral data by default. According to the official Claude Code settings documentation, the following are active unless you explicitly disable them: auto-memory (the system builds a persistent behavioral model of how you work), away summaries (session context is summarized and stored), and feedback surveys. Prompt history is retained unless you set CLAUDE_CODE_SKIP_PROMPT_HISTORY=1. Each of these defaults is a data collection path. The aggregate picture is a detailed model of your engineering workflows, your codebases, your tool usage patterns, and your decision-making cadence — collected by the same company whose models you are paying to use, with no open-source version of those models for you to run yourself.
That last point matters: Anthropic has released zero open-weight or open-source models. Every Claude model is proprietary, API-only. If Anthropic changes pricing, changes terms of service, changes what data they collect, or simply decides your use case falls outside their acceptable use policy, your options are to comply or rebuild. There is no self-hosted fallback.
the open-source alternative is no longer a compromise
The argument for tolerating this situation used to be performance. Closed frontier models from Anthropic and OpenAI were measurably better on benchmarks that mattered to engineering teams. That gap is closing rapidly, and in some cases it has already closed.
Google released Gemma 4 on April 2, 2026 — a fully open-weight model family that runs on Ollama, llama.cpp, MLX on Apple Silicon, and even in-browser via WebGPU. No NVIDIA GPU required for the efficient variants. The 31B dense model scores 85.2% on MMLU Pro and 89.2% on AIME 2026. The 26B MoE variant runs with only 3.8 billion active parameters — meaning a machine you already own can serve it — and scores 82.6% on MMLU Pro. It has been pulled from Ollama over 12.8 million times in the two months since release.
Meta's Llama family, Mistral's open releases, and DeepSeek's models (which TechCrunch reported closing the gap with frontier closed models in April 2026) collectively represent a model ecosystem where performance no longer requires a proprietary API contract. The benchmark gap that justified vendor dependence is gone for most enterprise use cases.
the competitive picture is shifting
OpenAI's position in the agentic AI space — the domain most relevant to enterprise automation — is weakening. According to Ramp spend data reported by TechCrunch in May 2026, Anthropic overtook OpenAI in business customer share for the first time: Anthropic at 34.4%, OpenAI at 32.3%, with Anthropic growing approximately 26 percentage points in twelve months. Meanwhile, OpenAI's own agentic product rollout has been marked by the same gating and restrictions they publicly criticized in competitors — Sam Altman criticized Anthropic's model restrictions, then implemented identical ones on OpenAI's Cyber model weeks later.
Neither company is building infrastructure that gives you control. Both are building infrastructure designed to be the control layer between you and your AI capabilities.
the compute arms race has a real-world cost
The infrastructure underpinning these services is worth examining. Google committed $920 million per month to SpaceX for approximately 110,000 NVIDIA GPUs — overflow capacity, because even Google's own infrastructure cannot keep pace with the demand from its Gemini Enterprise platform. Alphabet committed over $180 billion in capital expenditure for 2026. Anthropic is paying over $1.25 billion per month for compute at a single facility.
These are not neutral numbers. They represent a physical resource extraction — land, water, power, silicon — being deployed at a scale that has generated significant community opposition. Environmental advocate Erin Brockovich launched a national data center impact mapping campaign in May 2026, receiving roughly 4,000 community submissions within a month. Her core criticism: permits are secured before communities are informed, and local officials often sign non-disclosure agreements before any public input is sought. The xAI Memphis data center — built to serve Grok — became a focal point for concerns about noise, water consumption, and rising utility costs for surrounding neighborhoods.
When you route your workloads through a large model provider API, you are a downstream participant in this resource consumption. When you run models on your own infrastructure, you make a different choice.
the hardware ecosystem is opening
The NVIDIA dependency that made large model inference expensive is softening. The Gemma 4 26B MoE variant with 3.8 billion active parameters runs on consumer and workstation hardware via llama.cpp on CPU, or MLX on Apple Silicon, without requiring a datacenter GPU. AI chip startup Groq, now reportedly raising $650 million following a $20 billion licensing arrangement with NVIDIA, has pivoted toward inference-as-a-service — offering sub-100ms latency on open-weight models as an alternative inference path. The model execution layer is becoming a commodity, not a moat.
what agenticwork and our open-source core actually solve
The problem described above is not a model quality problem — frontier open models are competitive. It is not a tooling quality problem — the MCP ecosystem and agent frameworks are mature. It is a control and dependency problem.
Our platform and the open-source core are built on an explicit design premise: your model provider choice should be a configuration decision, not an architectural commitment. The platform's canonical map/reduce streaming layer means any model — Claude Opus 4.8, Gemini 2.5 Pro, Gemma 4 running on Ollama on your own hardware, or a fine-tuned DeepSeek variant — produces identical, consistent results through the same interface. Switching providers is a settings change, not a rebuild.
Specifically, this means:
- No token session limits if you are running your own models. Ollama on your own hardware means no per-token billing, no session caps, no rate limits imposed by a third party. The Gemma 4 26B MoE runs on a machine with 32GB of unified memory — hardware that enterprise teams already own.
- Your behavioral data stays yours. The platform runs in your environment. Auto-memory, session transcripts, prompt history — stored and controlled by you, not indexed by a model provider to improve their next release.
- Provider pricing leverage. When switching providers requires no code change, you can negotiate. You can A/B test. You can route cost-sensitive workloads to cheaper or open models while reserving proprietary frontier models for tasks that justify the expense.
- No dependency on a provider's acceptable use policy. Your agents operate tools, access cloud infrastructure, and take autonomous actions. Those actions should be governed by your policies, not a third party's interpretation of what constitutes appropriate use.
Our open-source core — self-hosted, one-command install, with an open-source release planned — is the clearest expression of this philosophy. Sixteen pre-wired MCP servers, Chat, Flows, and Code Mode, running on your infrastructure, against your providers, with no telemetry leaving your environment unless you configure it to.
the window for decisions is now
The teams that are building workflow automation, agentic pipelines, and internal tooling on top of proprietary APIs are accumulating a dependency that will be priced against them. The Tokenpocalypse is not a prediction — it is already documented in GitHub Copilot billing complaints and enterprise procurement discussions happening right now.
The open model ecosystem is ready. The inference hardware story has changed. The design pattern for provider-agnostic agentic infrastructure exists and is built into our platform and the open-source core today.
The question is not whether to build provider-agnostic — it is whether to start before or after the next pricing restructure lands in your budget cycle.
Trenton White is Founder & CEO of AgenticWork. Our platform is the enterprise-tier layer built on the open-source core, the open-source agentic platform for teams that want to own their AI stack.