// AI Deep Dive
Rent the Frontier, Own the Rest
Token prices keep falling and AI bills keep rising, and most of that money lands with two vendors. The fix isn't to spend less on AI. It's to stop renting all of it.
// Executive Summary
The story most companies tell themselves about AI costs is that the market will fix them. Token prices have collapsed, so the bill should shrink on its own. That isn't what's happening. Usage is growing faster than prices fall, and the money is concentrating with a handful of frontier labs and the chip and cloud vendors behind them. Meanwhile, open models have closed most of the capability gap, and the hardware to run them now sits on a desk. The question for 2027 isn't how much to spend on AI. It's how much of your intelligence you want to rent.
The AI bill is now a usage bill, and it's enormous. Gartner puts worldwide AI spending at $2.67 trillion in 2026, up 49.5%, and says inference overtook training this year in AI-optimized cloud infrastructure, $23.3 billion to $19 billion.
Most of the money goes to two vendors. Anthropic and OpenAI together take 67% of enterprise LLM API spend, by Menlo Ventures' estimate, and Anthropic's annualized revenue hit $65 billion in July, seven times a year earlier.
The open models already do most of the work. Open-weight models, whose files anyone can download and run, now carry 72% of the tokens on OpenRouter, a marketplace that routes traffic to hundreds of models, while closed models still collect 73% of the dollars. Mozilla's analysis puts the capability gap at about three points on the most-watched benchmark.
Rent frontier models for the work that needs them, run open models on hardware you control for everything else, and keep your data on the side of that line you own.
// The Hook
A Cannon by the Shot
So the cheapest AI I use all day runs on a Mac Mini in the corner of my office. It's a seven-billion-parameter version of Qwen, an open model from Alibaba, running through Ollama, and it costs me nothing past the hardware. Hermes Agent, the open source assistant that runs my morning routine, goes to that little model first. Summarize the email. Sort out what needs a reply. Rename the files. Draft the first pass of a social post. When it hits something the little model can't handle, it reaches out through OpenRouter and rents a frontier model for exactly that one job. I'm just hanging out in my office and everybody else is doing the work.
I didn't start there. For most of the last year I did what everybody did, which was point everything at the best model money could buy. Claude for the big stuff, ChatGPT for the rest, and a growing stack of subscriptions and API keys to keep both fed.
Then I actually looked at what I was asking the best model in the world to do. Summarize this thread. Pull the dates out of this PDF. Give me three subject lines. Mosquito work. I mean, it's like looking up who won the 1971 World Series by reading the whole Library of Congress. And I was paying frontier rates, by the token, for every swat. There's an old proverb that you shouldn't use a cannon to kill a mosquito. I'd been renting a cannon by the shot.
What changed my mind wasn't ideology. It was the meter. Once the routine work moved onto the Mac Mini, my frontier spending shrank to the jobs that actually deserved it, and those jobs got better because I'd stopped rationing. People are sleeping on small models. The bigger version of that story is playing out in enterprise budgets right now, and the numbers are a lot larger than mine.
// The Deep Dive
Rent the Frontier, Own the Rest
Falling token prices were supposed to make AI cheap. Instead they made it everywhere, and the bill flows to whoever owns the model.
The conventional wisdom says AI costs will take care of themselves. Mozilla's State of Open Source AI report notes that running a model with the performance of the 2023 frontier cost about $20 per million tokens three years ago and about 40 cents today. On that logic, you rent the best model available and never think about infrastructure again.
The complication is that cheaper units haven't produced smaller bills. Stanford's AI Index found the cost of querying a model at a fixed late-2022 quality level fell more than 280-fold in under two years, and over roughly the same stretch enterprise spending on generative AI more than tripled, by Menlo Ventures' count. Agents that plan, retry and call tools consume tokens by the thousand per task where a chatbot used dozens. And every one of those tokens is rented. The companies that treat AI as a pure utility aren't just paying a usage bill. They're shipping the margin on their own AI work to the few vendors who own the models, and sending their data along with it.
What the Intelligence Rent Is
The intelligence rent is the share of your AI value that leaves the building every month because you own none of the stack that produces it: not the model, not the hardware, not the place your data gets processed. Fine for frontier work. A problem when it's the default for all work. It shows up at three levels.
Individual: For a person, it shows up as a meter that never stops. Salesforce CEO Marc Benioff said his company expects to spend about $300 million on Anthropic tokens this year, mostly on coding, and that it's building systems to route tasks between large and small models depending on complexity and cost.
Team: For a team, it shows up as concentration. Menlo Ventures estimates that enterprise spending on generative AI reached $37 billion in 2025, more than triple the year before, and that Anthropic at 40% and OpenAI at 27% now take two-thirds of enterprise API dollars. Two vendors set the price, the terms and the roadmap for most of the work your teams depend on.
Organizational: For the enterprise, it shows up in the capital account. Gartner puts worldwide AI spending at $2.67 trillion this year, 56% of it infrastructure, and says that in AI-optimized cloud infrastructure inference now outspends training, $23.3 billion to $19 billion, with inference headed to 59% by 2027. The money is in running models, not building them, and running is the part you could own.
The labs you rent from are renting too. PitchBook reports that Anthropic's compute costs ran 71 cents of every revenue dollar in the first quarter of 2026, projected to fall to 56 cents in the second, on a gross margin around 44%. Your token dollar passes through the lab to a hyperscaler and on to NVIDIA. Nobody in that chain owns the intelligence except the chip, and the chip is the one thing you can buy.
How It Works in Business Contexts
It plays out in four stages.
Stage one: rent the best, measure nothing. Pilots run on the frontier API because it's the fastest way to see what works. Nobody asks which model a task actually needs. Enterprises moved further in this direction last year: open models' share of enterprise API spend fell to 11% from 19%, by Menlo's count, largely because Meta's Llama stalled while the labs kept shipping.
Stage two: the usage bill lands. Agents arrive, token counts explode, and finance notices. Broadcom's survey of 1,800 senior IT leaders found cost has overtaken security as the top public cloud concern, and that 62% are very or extremely worried about generative and agentic AI infrastructure costs.
Stage three: the open models catch up, quietly. Open-weight models crossed half of all tokens on OpenRouter in the spring and carried 72% in the most recent week, with Chinese labs supplying 69%. Qwen alone has 942 million cumulative downloads on Hugging Face. Mozilla's analysis, drawing on a survey of 1,494 developers and 100 trillion tokens of routing data, puts the open-closed gap at about 3.3 points on Chatbot Arena, with open models matching on coding and general knowledge and lagging on long agentic tasks. Airbnb runs Qwen for user-facing features, and Cursor built its internal model on Qwen.
Stage four: the work comes home. Broadcom found 56% of enterprises running or planning production inference on private cloud against 41% on public cloud, a reversal from a year earlier, and 43% of companies repatriating workloads are moving AI training and inference specifically. Cloudian's smaller survey of 203 IT decision-makers found 91% would choose on-premises, private or hybrid infrastructure for AI that touches sensitive data. Both surveys were vendor-commissioned, so read them as direction, not precision.
The hardware caught up too. NVIDIA's DGX Spark, a desktop box with 128 gigabytes of memory shared between processor and GPU, runs the open flagships locally, including DeepSeek, Qwen, OpenAI's open-weight gpt-oss and NVIDIA's own Nemotron family. It listed at $3,999 at launch, and NVIDIA raised the price to $4,699 in February, citing memory supply; that's the list price as of October 15. JetBrains' chief executive, quoted in the same NVIDIA post, described the buyers as organizations that need full control over security, governance and intellectual property. NVIDIA, which gets paid whether you rent or own, released a full family of open models sized from a laptop up to a data center. The largest open models, like Moonshot's Kimi at 2.8 trillion parameters with a recommended 64 or more accelerators, get rented from open-model hosts like Groq and Fireworks AI rather than run on a workstation.
Tier | What you pay (list prices, October 15, 2026) | What you control | Where it fits |
|---|---|---|---|
Rent the frontier (Claude, OpenAI, by API) | Your prompt and nothing else. The vendor sets price, terms, roadmap and data policy | Novel reasoning, long agentic runs, the slice of work where the best model changes the outcome | |
Rent an open model (DeepSeek, Qwen, Kimi, GLM through OpenRouter, Groq or Fireworks AI) | $0.15 to $2 per million tokens; DeepSeek's flagship lists at $0.66 in and $1.98 out off-peak | The model, which you can move to any host. Not the hardware or the data path | High-volume routine work, and cost-per-task testing before you buy hardware |
Own the inference (Ollama or LM Studio on a workstation or DGX Spark; a GPU rack for teams) | Hardware up front; a DGX Spark lists at $4,699; marginal cost per token near zero once the box is busy | Model, hardware, data and uptime. Also the staff to run them | Steady, predictable volume, and anything involving data you wouldn't paste into a public chat window |
I've seen this cycle before. Back in the 90s, when I was doing support at an ISP, I'd grab the decommissioned Pentiums from outside the IT department and turn them into Linux servers, because the alternative was renting your operating system from Sun, HP or IBM, bundled with the box, at whatever price they named. Linux let companies own that layer, and Red Hat built a very good business on the idea that you rent the help and own the code. Then cloud came along and re-rented the whole stack, and I was an executive at Cloud.com selling exactly that pitch. The correction is already visible outside AI: 37signals saved about $2 million in its first full year off AWS, on roughly $700,000 of hardware, against a cloud bill that had reached $3.2 million. AI is running the same loop, just faster, and the open models are the Linux of this cycle.
How to Build a Rent-or-Own Strategy
Phase 1: Sort the mosquitoes from the cannons (the first 30 days)
Most AI work is routine. Find out how much of yours is.
Pull 30 days of usage from each frontier account and classify the tasks: summarize, extract, classify, draft and translate on one side; novel analysis, long multistep agent runs and customer-facing reasoning on the other.
Tag each production workflow with the data it touches. Anything with customer records, source code or contracts goes on the "own" list by default.
Pick three high-volume routine workflows as candidates to move.
Phase 2: Rent an open model and measure cost per task (days 31 to 60)
Price per token isn't price per task. Test before you buy hardware.
Run the three candidates on an open model through OpenRouter, Groq or Fireworks AI alongside the frontier model, and compare cost and quality per completed task, not per token. Some open models burn far more tokens to finish the same job.
Keep the frontier model as the fallback for the cases the open model fails, and log how often that happens. That ratio is your rent-to-own split.
Read the license on every model you adopt. DeepSeek ships under MIT; Kimi and Nemotron ship under their makers' own terms.
Phase 3: Own the inference for steady, sensitive volume (days 61 to 90)
Buy hardware for work that's predictable and private, not for peaks.
Start with a workstation-class box, whether a DGX Spark, a Mac Studio or a single GPU server, running Ollama or a production serving stack, loaded with the open model that won Phase 2.
Route by policy, not by person: local first, open-model host second, frontier last, with the agent choosing by task type and every escalation logged.
Hold the frontier contract for what's left, and renegotiate it on the smaller volume.
The companies getting this right share a few things:
One routing layer. Every agent and app goes through it, so the decision is made once, in policy.
Cost per completed task. Measured on their own workload before any hardware purchase.
Utilization targets for owned hardware. An idle GPU is the most expensive AI you can buy.
A named owner for licenses and data residency.
Common Missteps
Comparing price per token instead of price per task. A model that costs a tenth as much per token and uses five times as many to finish is only half as cheap. Mozilla's report also notes that fewer teams get open models into production, 51% against 63% for closed models, mostly on tooling and trust rather than capability. Measure the finished job.
Buying hardware for the peak. Owned GPUs pay off when they're busy. A rack sized for your busiest hour sits idle the rest of the week. Size for steady volume and rent the peaks.
Treating "open" as one thing. Qwen's small models run on a laptop; Kimi's flagship wants 64 accelerators. DeepSeek ships under MIT; Nemotron ships under NVIDIA's own license. Policy risk is real too: most of the open frontier is Chinese, and your general counsel belongs in that conversation early.
Owning the box but not the data path. Running a local model while your agents still ship the full context to a hosted tool is letting the fox in the henhouse through the side door. Draw the line around the data first, then put the compute on the right side of it.
Business Value
ROI considerations:
Moving routine, high-volume work to open models cuts the per-token cost by an order of magnitude at current list prices, and that's where most of the volume lives.
Owned inference turns a variable usage bill into a fixed asset with near-zero marginal cost, the same math 37signals used to recoup its hardware inside a year.
A routing layer gives you real negotiating power with frontier vendors, because you can show how much of your volume could leave.
Competitive implications: The companies that rent everything will keep paying a growing share of their AI value to two labs and their suppliers, and they'll live with those vendors' pricing, outages, export rulings and roadmap changes. The ones that split the stack keep the frontier where it matters, own the rest, and keep their data. In a market where 72% of tokens already run on open models, the second group is buying what everyone else will be buying in two years, at this year's prices.
// Key Takeaways
Stop treating the best model as the default. Most daily AI work is routine. At frontier list prices of up to $50 per million output tokens, routing a summary or a classification to the top model is paying cannon rates for mosquitoes.
Follow your AI dollar to where it ends up. Two vendors take 67% of enterprise API spend, and Anthropic's own compute costs ran 56 to 71 cents of every revenue dollar this year. Renting everything ships the margin on your AI work up that chain.
Rent open models before you own them. Open-weight models already carry 72% of OpenRouter tokens at a fraction of frontier prices. Test cost per completed task on your own workload through an open-model host before buying hardware.
Own the inference for steady, sensitive volume. With 56% of enterprises moving production inference to private infrastructure and a desktop box now running the open flagships, the data you'd never paste into a public chat window belongs on hardware you control.
// What This Means for Your Planning
So in planning terms, the AI budget has two parts now, and most companies are budgeting only one of them. There's the frontier rent, which should stay variable and tied to the work that needs the best model. And there's owned intelligence, which is capital: hardware, open models and the people who run them. Budgeting all of it as rent is like leasing every car in the fleet forever because the dealer keeps cutting the monthly rate.
The boardroom assumption I'd push back on is that renting is the low-risk choice. I get why it feels that way. But a company whose entire AI capability sits behind two vendors' APIs carries price risk, availability risk and policy risk it doesn't control. Anthropic had to switch off its two most advanced models for about two weeks in June to comply with an export directive.
For the 2027 cycle, I'd do three things. Classify the AI work before you fund it, so the frontier line covers frontier work and nothing else. Fund a 90-day open-model test with a cost-per-task number as the deliverable. And put a hardware line in the budget for the steady, private volume, sized for what runs every day rather than what might run on the busiest one.
Look, we all want to be on the frontier when the frontier is where the work is. The rest of the time, use the tool that fits the job. Here's the question for your next planning meeting: of everything we asked AI to do last month, how much actually needed the best model in the world, and what did we pay for the part that didn't?
Your AI Sherpa,
Mark R. Hinkle
Founding Publisher, The AIE Network
Follow me on LinkedIn
If you want to get in contact or give me feedback, reply to this email. I read every single one of them.
