In 2004, Tiger Woods was already the greatest golfer alive. Eight majors. Years as world number one. Nothing in the data said fix anything. He rebuilt his swing anyway. Working with Hank Haney, he dismantled the mechanics that had produced his greatest results and started over from the ground up. Hi s driving accuracy fell to 182nd. The press said his game was coming apart. Three years later he had the best season of his career. That is the dilemma almost every company we work with is in. The business is running. The numbers are fine. Someone on the team is tweaking a workflow with a language model and calling it AI adoption. The easy move is to leave it there and renew the SaaS contracts for another year. It is also the biggest risk sitting on the balance sheet. Becoming AI-first requires something closer to Tiger’s rebuild. Thirty or fifty specialized models, orchestrated into a recursive organization that gets smarter every quarter. I made the full architectural case in Infinite Leverage. Renewing with a legacy SaaS vendor who glued an AI wrapper on last year’s product is shuffling deck chairs on the Titanic, which I covered in The AI Shuffle. Language models are a good place to start. They are not a place to stop, and the frontier labs want you to. I wrote in Peak Token that the frontier model makers were pricing themselves out of their own market. The capability gap between the best closed model and the best open one has nearly closed while the price gap stayed enormous. A commodity priced like a luxury does not stay a luxury for long. The frontier labs see the same data. The response is the standard reflex of a software business whose moat is eroding: hold customers on inertia when you can no longer hold them on capability. I covered the broader software collapse in Software Is Not Going Down Alone and the incumbent debt response in The SaaS Debt Trap. Which brings me to the word I want leaders to start using. Choke points. A choke point is any place in your architecture where a single party decides what you pay, what you can use, and how fast you can move. Every stack has them. Some are wonderful. A few are worth running toward. Most will quietly cost you your freedom. The most dangerous one is the layer almost nobody in the boardroom is discussing yet. The one that made everyone richer History as context. The best network effect I ever built was at LinkShare. We dominated affiliate marketing, and the dominance was good for everyone who touched it. For the affiliates sending traffic to merchants, our scale was a gift. One place to find the best merchants. One set of reports. One check at the end of the month instead of chasing fifty small payments. For the merchants, the consolidation was equally valuable. One network with the scale to hunt scammers gaming commissions and the plugins quietly hijacking commissions at checkout. A thousand fragmented programs could never have won those fights. One network could. Nobody at LinkShare was trapped. Affiliates and merchants stayed because leaving meant giving up the thing that made their lives easier. The network and the customer’s interest pointed the same direction. That is the shape of the best businesses ever built. Google search improves as more people use it. Waze gets more accurate with every driver. eBay is worth more to a seller because the buyers are there, and worth more to a buyer because the sellers are there. PayPal got safer the more it connected. Collective[i] gets smarter about economic behavior with every new company on the network. Intelligence.com maps professional relationships, and every new member sharpens the map. The door is unlocked the whole time. The network itself is the reason to stay. The one that holds you hostage Now the other kind, the second-order effect from decisions you make today. Too many leaders are not thinking about what it means to be AI first. The AI labs are trying to take advantage of this by learning from software. SQL is a public standard. Any database runs it. Oracle invented PL/SQL as a gift. Bundle your logic into the database, run it faster, do things plain SQL could not. The other half of the deal was that every procedure got written in a language only Oracle spoke. The standard SQL stayed portable. The PL/SQL wrapped around it did not. Over years, decades of business logic ended up fluent in a dialect no other database understood. By the time a company wanted to leave Oracle, the bill was not a software license. It was a rewrite of decades of code. Microsoft ran the same play against WordPerfect: easy to bring your old files in, impossible to take them cleanly back out. The format itself became the choke point. What keeps you there is not the value of the product. It is the cost of leaving. When that cost exceeds the value of staying, you are no longer a customer. You are a hostage with a renewal date. The gap between the two lines is the margin the vendor took from you. I wrote about the walled version of this play in Karp and Nadella Are Selling You a Wall. In AI, lock-in moves overnight Software lock-in played out over years. Executives had time to spot the trap and plan the rebuild. AI does not grant that courtesy. The enterprise mood shifted off ChatGPT and onto Claude in a matter of weeks after the Department of War contract decision. Then a second wave moved to open-weight models shortly after and in rapid fashion. If the most sophisticated software businesses on earth could not control the switching when it hit them, do not assume you will control yours. You are going to need many models in the next year, not one, and the specialized ones will not come from OpenAI or Anthropic. Locking into a single language-model vendor today locks you out of twenty-nine of the thirty most important models you will deploy. That is not a cost problem. It is an existential architecture mistake you will spend the entire next cycle paying for. At Collective[i] we live this cycle every week. The model mix underneath Telli Assistants shifts constantly. When a new open-weight release lands, we test it against the sales workflows that matter and promote the winner within days. When a frontier release gets quietly worse at a specific job, which happens, the harness routes around it the same hour we see the drift. The customer never sees the swap, which is the point. The companies running their most important workloads on a single-vendor stack are running them on last Tuesday’s model, hoping the price does not move and the capability does not drift. Hope is not an architecture. A moat by any other name None of this makes the vendors villains. The founders and investors behind the frontier AI labs learned their craft in software, and they are applying the software playbook to AI with precision. Investors do not fund software on revenue alone. They fund it on defensibility. Warren Buffett built a career on the word moat. Building barriers to exit is not a deviation from the software business. It is the software business, imported directly into AI. What changed is that the old moats stopped working. A Morningstar analysis in early 2026 found that four of the five classic moat pillars had lost most of their predictive power under AI exposure, and the companies most exposed to AI underperformed the resilient ones by close to 26 percentage points in the first months of the year.1 I covered the mechanics in The Arc of a Company’s Life. The frontier labs read the same data. They know the capability gap is closing and the price gap is a factor of ten. Their response has two moves. The first move is regulatory capture. I made this case at length in The Death of Humanity. The frontier labs are pushing for AI regulation framed as safety. The regulation they want happens to be exactly the regulation that would knock the open-weight competition out of the enterprise market. Make the compliance bar high enough and only the labs with the money and lawyers to clear it get to serve the Fortune 500. The call for safety becomes the moat the capability no longer provides. They are not raising the alarm because they fear a smarter model. They are raising the alarm because they fear a free one. The second move is forward-deployed engineers. The labs are sending their best engineers inside the Fortune 500 to build the enterprise’s AI integration for them, at zero cost. On the surface it is a gift. In practice every production workflow ends up wired to that lab’s model, that lab’s conventions, that lab’s assumptions. The engineer rotates out after six months. The dependencies stay. The lab did not accidentally build a hook. They built it on your dime, and you thanked them for it. The strategic logic sits one level up. These labs know they will not build every model your business will run. They know the thirty-brain company is coming. What they want is to be the gatekeeper every other model has to pass through to reach you. If the orchestration layer belongs to one lab, every other model vendor becomes their subcontractor, not your supplier. The day you cede that layer is the day you turned your business into a subscription to one lab’s roadmap. Six places a choke point can sit in your AI stack Point that distinction at AI and you find a choke point at every layer. They are not the same kind, and sorting them is the heart of the architecture decision. Six places sit in the AI stack, sorted by who they actually serve. The network at the top is the one you want. The harness, memory, data, and system-of-record layers are the ones a vendor uses to hold you. The model in the middle is fine to lean on while it is best. That window is closing fast. 1. The model. The commodity choke point. The weakest layer the frontier labs are standing on. Everyone has roughly the same thing. So the labs convert the model into a choke point by getting your internal data inside their walls and by making themselves the answer to your governance and compliance worries. Governance is often where lock-in enters wearing a responsible suit. I made that case in The Safest Move Will Cost You Everything and in the AI-First Framework for Responsible AI. Each new open release narrows the gap. DeepSeek V4-Pro landed mid-2026 at 80.6 percent on SWE-bench Verified, within a quarter of a point of Claude Opus 4.6 at 80.8. Kimi K2.6 passed Opus a quarter later. Kimi K3 is the one leading the pack as I write this. The old twenty-to-thirty-point knowledge-benchmark gap has shrunk to three to five, and shrinks again at every open release. GPT-5.5 output lists at roughly thirty dollars per million tokens. An equivalent open-weight API runs around four. You are paying luxury rates for commodity work.2 Grok Bot shipped on August 11, 2026. Five weeks later it was on roughly a million devices and running at a thirty-seven-million-dollar monthly mobile revenue clip, more than double what xAI’s mobile line brought in nine months earlier.3 A single consumer product on a non-frontier model moved the whole business line in one quarter. Abstract this layer away, log every call, swap at will. 2. Memory and context. The quietest hook. Grows every day and invisible. Your agent’s memory, accumulated context, embeddings, prompts refined over months, all held in a format only the vendor can read. The most PL/SQL thing in the stack. The vendor calls it a feature, and it is, right up until the day you leave and discover that everything the system learned about your business stays behind. 3. Data gravity. Leaving means leaving what you taught it. Once your proprietary data lives inside one provider’s fine-tuning pipeline, switching means walking away from the thing you spent years teaching. Vendors sell this hard as a benefit. The model knows your business. It does. That is exactly why it is a hook. 4. The system of record. When the AI becomes the desk. The Salesforce move, applied to AI. When the assistant becomes the place your people work all day, ripping it out stops being a swap and becomes a company-wide event. The SaaS incumbents know they are being ripped out, and they do not have a coherent answer. Some are opening the doors with headless APIs so they can stay the data layer under whatever AI front end wins. Salesforce is the clearest example. Others are quietly welding the doors shut. Slack is blocking enterprises from using their own Slack data to train their own models, at the same moment the parent company is pitching open AI partnerships on the other side of the house. Same company, two opposite strategies, zero plan. The move that works is to pull your data out of the SaaS system of record and into a store you own, usually Snowflake or Databricks. That strips the vendor of their last hold over your AI substrate. The SaaS giants are borrowing billions to buy back their own stock in the quarter they should be rebuilding their products on an AI substrate. They are not. That is why everyone else is building the AI substrate around them. 5. The network. The choke point worth wanting. The natural one. The LinkShare pattern. The Collective[i] pattern. The Intelligence.com pattern. Gets better for everyone as it grows, because every new participant widens the signal the whole network learns from. The only layer where the vendor’s incentive and the customer’s incentive point the same direction. I come back to it at the close. 6. The harness. The one that decides everything else. The layer almost nobody is watching. The one that controls all the others. The harness is the choke point that matters The harness is the most important thing in AI, and the least talked about. It is the orchestration layer that turns many models into one business outcome. It is where recursive learning lives. It is where the logic of your company sits, not inside the model but above it. It is where every dollar of value AI produces for your business either accrues to you or walks out the door with someone else. Build it in the wrong place and your company ends up running on whichever models your provider has decided to promote this quarter, at whichever price they decide to charge next. Cede it and that someone else becomes the gatekeeper to every outcome your company can produce. Every model call and every agent decision passes through this layer. Every lesson your agents learn gets written down here or nowhere. Every token you burn gets priced here. Own it and you set the terms. Rent it and someone else does, forever. Claude Code is an excellent product and a proprietary harness, tightly bound to one model family. You cannot swap the model. You cannot fork the harness. Independent 2026 harness comparisons rank it strongest on code execution and weakest of the major harnesses on multi-model flexibility.4 The design is working exactly as intended, and the intent is to make the model and the harness one inseparable product. Oracle did the same thing with PL/SQL thirty years ago. The open harnesses are built on the opposite premise. OpenClaw, Peter Steinberger’s MIT-licensed project, carries the biggest community footprint in the category at roughly 354,000 GitHub stars and the broadest configurability. Hermes takes a research-first, self-improving approach and leads on memory architecture. OpenCode, Aider, and Pi share the one property that matters. The harness code is readable and forkable. You can run it on your own infrastructure. The difference between these and a vendor’s coding agent is the difference between owning the machine and renting time on someone else’s. Many models, working in concert Assume from the start that you will run many models at once. Not several flavors of chatbot. A working collection of different kinds of intelligence, each built for a different job. A language model to read and write. A reasoning model to work through a hard problem step by step. An economic model to forecast what a market or a buyer is likely to do, which cannot be faked by a chatbot with a confident tone. A robotics or world model for physical space. Vision, speech, and video models for the modalities the work arrives in. Small local models for high-volume repetitive tasks. Retrieval and embedding models for context. Each is a specialist, and the good ones are not interchangeable, any more than a cardiologist is interchangeable with a structural engineer. I made the full architectural case in the AI-First Framework for Responsible AI. Your company will run on dozens of these specialists over the next few years. No single frontier lab will build all of them. The frontier labs are good at the language and reasoning columns. They are not building the economic forecasting brain, the physical-space robotics brain, or the specialized small models that will do most of the actual work. Locking yourself to Anthropic or OpenAI is a bet that one lab will deliver every brain your company needs. It will not. Language models are the single most commoditized piece of the AI world. If you have ever heard a CEO say “we have AI” and the proof is that their teams use ChatGPT or Claude, you are talking to a company that mistook the entry drug for the operating system. I made this architectural argument in The Next Computer Is Alive. All AI is narrow intelligence. The winning move is many specialists in concert, not one generalist asked to do everything. The recursive loop lives in the harness Agents that improve themselves are no longer a thought experiment. Karpathy’s AutoResearch ran seven hundred ML experiments on one GPU over two days with no human in the inner loop and found twenty ways to make training faster. Self-Harness has agents improving their own harness through a propose, evaluate, accept loop. DeepMind’s AlphaEvolve found a better matrix multiplication algorithm the same way.5 The pattern is the same every time. An agent acts, the work gets scored, the lesson gets written down, and the next run is better. Recursion exists, and the moat moved to the harness. The loop that makes your agents smarter every week does not live inside the model. It lives in the layer that scores the work, stores the lesson, and feeds it back. That layer is the harness, and its memory needs to be auditable, portable, and yours. I covered the organizational version in Infinite Leverage and The Agent Swarm Just Killed Scale as a Moat. Rent the harness and you have rented the loop. Every lesson your agents learn accrues to someone else’s proprietary format, and the compounding intelligence you thought you were building is sitting in their warehouse with your name on none of it. The harness sets your costs Picture a manufacturing machine that dictates the price of every pipe you are allowed to buy, when the same pipe sells elsewhere for a fraction. The machine will only run on its own pipe. That is a coding agent bound to one model family. Uber and most of the heavy token users have already learned that agentic workflows blow through budgets without warning.6 Lose control of the harness and you lose control of the bill, permanently. Routing is where the money is Once you own the harness, the cost problem becomes solvable. OpenRouter gives you one API across hundreds of models and picks a capable one for each query. Rob May, who founded Backupify, co-founded Neurometric.ai in January 2025 for task-based routing. His customer arc is always the same. A company burning two hundred thousand a month on frontier inference identifies its top workload, routes it to a small specialized model, and watches the line collapse. Then they do it again with the next workload. Frontier usage does not go to zero. It goes to where a frontier model was actually needed, which is a much smaller share than anyone assumed. Image, audio, and video are commoditizing on the same curve. A harness that can route across all four modalities, choosing the cheapest model that clears the quality bar for that specific job, is the difference between an AI budget that scales with your ambition and one that scales with your vendor’s pricing page. Rent the model. Own what compounds. The whole thing collapses into one decision made over and over. What to rent, and what to own. Rent the parts that are becoming a commodity and keep them swappable. Own the parts that compound and would be expensive to rebuild, and keep them portable. The line between the two columns is the entire strategy. Rent the model. It is the fastest-commoditizing part of the stack. Treat raw inference as a spot market. Own three things, and keep all three portable. First, the gateway, so switching models is a configuration change with a rollback rather than a rewrite. By April 2026, 81 percent of enterprises were running three or more model families in testing or production, up from 68 percent a year earlier.7 The single-vendor era lasted eighteen months. Second, the agent blueprint. Hold your prompts, tool definitions, retrieval rules, and settings as provider-neutral configuration, with verification tests that prove behavior still works after a swap. Lean on the open standards forming around exactly this, the Model Context Protocol chief among them, now stewarded by the Agentic AI Foundation that Anthropic, OpenAI, and Block helped launch. Third, and most valuable, the context layer. Context engineering has replaced prompt engineering as the core discipline. The hard part is deciding what the system should know, how fresh that knowledge must be, who is allowed to see it, and what each retrieval costs. You can rent a model. You cannot rent the knowledge architecture built around your own operations. We built Telli Assistants on this architecture from day one. A harness that lets a company connect any model and any dataset and compose any agent the business needs. We assumed enterprise customers wanted the freedom to swap models in and out, because a network-effects company has no reason to lock its users to someone else’s pricing page. We started in GTM and relationships because those are the arteries that touch every function. The same architecture serves the private equity and venture capital firms on the network. The one choice we made for them was to keep the harness open. What getting this wrong costs When Builder.ai collapsed, one manufacturer spent more than three hundred thousand dollars and three months migrating forty AI workflows to a new platform, with customer-facing features degraded the entire time, all of it avoidable with a provider abstraction in place from the start.8 The quieter cost is larger. A provider under no pressure to compete has no reason to pass along the price collapse happening everywhere else. You will pay luxury rates on commodity work, forever, and the number only surfaces when someone asks why the AI line tripled. The upside is the mirror image. Rent the model, own the harness, own the context. Route every job to the cheapest capable model. Swap a stronger model in the day it ships. Treat a provider outage as a failover. Compound your own intelligence in a layer that gets more valuable every quarter, regardless of which model is underneath. Two kinds of data advantage Which brings me back to the one choke point worth running toward. If proprietary data is the one thing a competitor cannot buy, there are two very different ways to turn it into an advantage, and most companies build the wrong one. The first is the data well. You feed your internal data into your own model and it improves, for you, on the narrow slice of the world your company can see. Researchers call this within-user learning. It mostly produces switching cost rather than real advantage. A deeper well, dug alone. The water is real, the well only serves the one who dug it, and the main thing it builds is the cost of climbing back out. Internal data as a moat is a polite way of locking yourself in. The second is the network refinery. A network learns across the whole market, and every member’s activity sharpens the intelligence delivered back to every other member. The difference is the difference between studying your own four walls and studying the external world. That is how Collective[i] and Intelligence.com are built. Network refineries. The more companies on the network, the sharper the insight each member receives, because the system is learning from the outside world. I made the deal-process version in The Buyer Has a Process and the Intelligence.com version in The Warm Intro Is Dead. The value compounds for everyone in the network. The door is unlocked. The reason to stay is the network itself. It is also the only data advantage that gets stronger as the models underneath become a commodity, because it never depended on the model. Flexibility is the only strategy that survives The whole argument fits in one line. Rent the model, own the harness, own your context, and join the network that refines the outside world rather than digging the well that only deepens your own. The reason this matters so much right now is that we are early. The best model today will not be the best model in six months. The harness you choose will need to run models that have not shipped yet. Every architecture decision made on the assumption that today’s winner stays the winner is a decision that ages badly. Flexibility and testability are the two properties that matter more than any other. The companies that build this way get to run experiments. The companies that lock themselves to one provider’s stack get to have opinions. That distinction is the seed of the next piece. The deepest cultural divide I see between companies winning with AI and companies performing it. One organization runs on belief. The other runs on hypothesis. I started making that case in The Hypothesis Company and will finish it next. For now, go find your choke points. Notes Morningstar analysis of AI exposure and classic economic moat pillars, early 2026. The four pillars that lost most of their predictive power under AI exposure: switching costs, network effects, intangible assets, efficient scale. The remaining pillar, cost advantage, showed partial resilience. Companies most exposed to AI underperformed the most AI-resilient cohort by approximately 26 percentage points in the measured window. Open-weight vs. closed-model benchmark comparison, 2026: DeepSeek V4-Pro at 80.6% on SWE-bench Verified vs. Claude Opus 4.6 at approximately 80.8%; Kimi K2.6 reported above Opus 4.6 on the same benchmark; Kimi K3 released fall 2026 as the newest in the series; Qwen 3 235B at 77.2% on GPQA Diamond. Pricing: approximately $30 per million output tokens for GPT-5.5-class closed-frontier models vs. approximately $4 for equivalent open-weight APIs. Knowledge-benchmark gap narrowed from 20-30 points in 2023 to 3-5 points in 2026. Appfigures, mobile app store revenue data. Grok Bot launched August 11, 2026. By September 19, 2026, approximately 1 million devices and approximately $1M in net mobile revenue. Grok and X combined earned $57M in August 2026 (Grok $37.4M, X $19.6M), up from $18.1M for Grok in January 2026. Independent 2026 harness comparisons across code-generation agents (Claude Code, OpenClaw, Hermes, OpenCode, Aider, Pi) ranked Claude Code highest on single-model code execution and lowest among the leading harnesses on multi-model flexibility and configurability. Andrej Karpathy, AutoResearch: a self-improving ML research loop running 700 experiments on one GPU over two days to identify training-speed improvements. Self-Harness: 2026 research on agents that iteratively improve their own harness through propose/evaluate/accept cycles. DeepMind AlphaEvolve: evolutionary code search that improved a matrix multiplication algorithm beating known baselines. Uber and other large enterprise AI consumers have publicly reported that agentic workflow spending can grow without warning when model usage is not routed and metered. a16z Enterprise AI Survey, April 2026. 81% of enterprises using three or more model families in testing or production, up from 68% in May 2025. The report characterized the single-vendor enterprise AI era as having lasted approximately eighteen months. Builder.ai collapsed in 2025. Reported downstream customer migration costs ran into the hundreds of thousands of dollars for enterprise customers that had wired Builder.ai agents directly into their operations without a provider abstraction layer. The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune. Messer is a retired lawyer, but none of this is legal, financial, or technical advice. Specific architecture decisions should be made with the engineers, counsel, and risk teams who know your situation. This story was originally featured on Fortune.com ...read more read less