This paper is written for the people who decide where enterprise AI actually runs. The CIO who owns the infrastructure budget, the platform leader who builds the stack, and the architect choosing between a model in someone else’s cloud and one on hardware they control. It answers one question. As hardware and models evolve through 2027, which of your workloads should stop being rented and start being owned.
The Local Inflection, the Hybrid Default, the Sovereignty Threshold, and the Inference Crossover are original constructs of this work, published under the license on the back cover. They build on measured research into local deployment under budget constraints [1], query routing across model tiers [2], and low-bit model efficiency [3]. Every quantitative claim is cited to a named source or labeled directional, without exception.
For three years the honest answer to where should your AI run was simple. In the cloud, because that was the only place a capable model would fit. You rented intelligence because owning it was not an option. That era is ending, and most plans have not noticed.
Three things moved at once. A capable open model now fits in a single machine you can buy. The same model, squeezed to a quarter of its original size, barely loses a step. And the smartest architecture is no longer local or cloud but both, with the easy work handled on your own hardware and only the hardest sent out. The result is that a growing share of enterprise AI can come home, where the data already lives.
This paper is my forecast for 2027, and the framework to act on it before the market agrees. Not a prediction that everything goes local, which is wrong, but a precise claim about which workloads cross, when, and why. The constructs are original work. The numbers are cited or labeled directional, without exception. I hold this to the standard of the rooms it will be read in.
The default of the last three years, rent your intelligence from the cloud, was correct when nothing else was possible. Through 2027 that stops being true for a widening set of workloads. The question shifts from whether you can own your intelligence to which parts you should. Five findings follow.
The rental default is breaking. Capable models only ran in hyperscale datacenters, so you rented. Three forces are now moving the line, and the moment they cross it is the Local Inflection. It arrives, workload by workload, through 2027.
It is not about cost. Cheap hosted models can undercut local on price alone. Local wins on data sovereignty, predictable latency, and control, and the economics only favor it at high, steady utilization. Anyone selling local on cost is selling the wrong thing.
The architecture is hybrid, not local. The Hybrid Default keeps the routine 70 to 90 percent of inference on your own hardware and routes the hardest 10 to 30 percent to a frontier model [2]. Owning everything is as wrong as owning nothing.
Hardware is the catalyst. Unified memory now puts a large model in one box, and 2027 silicon attacks the remaining limit, memory bandwidth. Paired with 4-bit models [3], the hardware is what moves the crossover, not a change of heart.
Decide with the Sovereignty Threshold. Four variables, data sensitivity, volume, latency, and task difficulty, tell you which workloads cross first. Map them now, pilot the hybrid, and move the ready ones home before the market calls it obvious.
Every enterprise AI stack built since the models got good made the same assumption. Capable intelligence lives in a hyperscale datacenter, so you call an API, you pay per token, and your data takes a trip outside the building on every request. That was not a choice. It was the only option, because a frontier-class model would not fit anywhere else.
The assumption is now expiring. A capable open model, squeezed to 4-bit, fits in roughly 35GB [3], well inside a single machine you can buy with 128GB of unified memory. The thing that forced you to rent, that the model would not fit, is no longer true for a growing list of workloads. The default did not become wrong. The ground under it moved.
When the model fits on hardware you own, renting it stops being a necessity and becomes a decision you have to justify.
Once the model fits, the question flips. Exhibit 2 shows the shape of that flip over time, and the point where it tips.
An inflection needs a cause. This one has three, and the reason 2027 matters is that they arrive at the same time. Any one of them alone would be interesting. Together they move the decision.
Unified memory now holds a large model in one machine. A 70B-class model runs on a box with 128GB, and designs reaching 256GB and beyond are arriving. Bandwidth is the last constraint, and 2027 silicon targets it.
4-bit is the shipping standard, and frontier open models now release in it [3]. A model at a quarter of its original size keeps almost all of its quality, so the thing that has to fit keeps shrinking.
You no longer pick local or cloud. A router sends the easy majority of work to a local model and the hard minority to a frontier one [2], so local does not have to win every case to win most of them.
The mistake on both sides is absolutism. Rent everything and you overpay and expose your data. Own everything and you buy frontier-grade hardware to do work a small model handles. The Hybrid Default keeps the routine majority local and routes only the hardest minority out.
| Workload | Runs on | Why |
|---|---|---|
| Retrieval, summarize, classify, extract | Local tier | High volume, shallow judgment, sensitive data. A small local model handles it well and keeps the data in-house. |
| Drafting and routine generation | Local tier | Frequent and forgiving. Quality is good enough at 4-bit, and the per-token cost of owning it is near zero at volume. |
| Hardest reasoning, novel problems | Frontier tier | Rare and demanding. Worth routing out to the best model, because the volume is small and the value per call is high. |
| Regulated or proprietary data, any task | Local tier | Sovereignty overrides difficulty. If the data cannot leave, the work runs where the data already is. |
There is a volume at which owning your inference costs less than renting it. It exists, and Exhibit 5 shows it. But leading with cost is a trap, because the line moves with utilization, and a cheap hosted model can sit below you anyway.
Owning inference is a fixed cost that you pay whether the machine is busy or idle, plus almost nothing per token. Renting is near zero fixed and a real cost on every call. So the crossover is not set by the hardware price. It is set by utilization. A box at 30 percent load loses to the cloud. The same box at 80 percent wins comfortably.
Which is why cost is the wrong headline. A cheap hosted open model can undercut your local box on price alone. What it cannot give you is the data never leaving the building, latency you control, and a capability that does not change under you. Those are the reasons to own. The economics only have to not get in the way.
The inflection is not a switch, it is a sequence. High-volume, shallow-judgment work on sensitive data crosses first, because a small local model handles it and the data wants to stay home. The hardest reasoning crosses last, or never. Knowing the order tells you what to move this year.
Owning your intelligence is not a platform rebuild. It is three phases on your first qualifying workload, each with a deliverable and a gate. Finish the phase or do not advance. In a quarter you can have a measured share of your inference running on hardware you control.
Four variables decide whether a workload should come home. Score each one, and the order falls out. Sovereignty can override the rest, and the hardest reasoning may never cross at all.
Must the data stay in-house for regulation, contract, or trade secret? If yes, this can override every other variable.
Is there enough steady work to clear the Inference Crossover, so an owned box runs busy rather than idle?
Does the workload need low, predictable response time that a round trip to a shared cloud cannot guarantee?
Can a small local model do the job at quality, or does it genuinely need a frontier model for the hardest cases?
The Hybrid Default keeps the routine 70 to 90 percent of inference local and routes the hardest 10 to 30 percent out. The Inference Crossover decides the economics, the Sovereignty Threshold decides the rest.
| Probe | Question | Yes | No |
|---|---|---|---|
| 1 MAP | Have you inventoried which workloads could run local and scored them? | ||
| 2 SENS | Do you know which workloads carry data that must stay in-house? | ||
| 3 FIT | Have you confirmed a 4-bit open model meets quality for your top workloads? | ||
| 4 BOX | Do you have, or can you buy, hardware that fits the model in one box? | ||
| 5 ROUTE | Is there a router that sends easy work local and hard work to a frontier model? | ||
| 6 UTIL | Will the owned hardware run at high, steady utilization rather than idle? | ||
| 7 LATENCY | Have you identified workloads needing latency the cloud cannot guarantee? | ||
| 8 PILOT | Have you run one workload end to end on the hybrid? | ||
| 9 SHARE | Do you measure and report the local share of your inference? | ||
| 10 FALLBACK | Does the hardest minority still route to a frontier model by rule? |
Your score places you on the posture model of Section 06. The posture you land on is the one to finish before you climb.
This paper is a conceptual and forward-looking contribution. The Local Inflection, the Hybrid Default, the Sovereignty Threshold, and the Inference Crossover are original constructs of this work. They rest on measured research into local deployment under budget constraints [1], query routing across model tiers [2], and low-bit model efficiency [3]. The hardware and economics figures are cited as dated, directional illustrations, because the specifics move quickly. The thesis is the trajectory, not any single number.
The 2027 call is a projection, built on forces already in motion and labeled as such. It extends an earlier paper of this series, The Future Is Local, which made the general case. This paper fixes the date, names the architecture, and gives the decision framework.
Arjun Jaggi, The Local Inflection, Enterprise AI Research, White Paper No. 09, arjunjaggi.com, November 2026.
Arjun Jaggi writes and advises at the intersection of enterprise AI strategy, model and infrastructure economics, and the operating work of deciding where AI actually runs. His frameworks turn the hardest questions of running AI at scale into terms a board can decide on and a team can execute. Engage at arjunjaggi.com.