Open Source LLM vs. ChatGPT Enterprise: Which Costs Less?

Is self-hosting an open source LLM cheaper than ChatGPT Enterprise?

Open-weight self-hosting beats a closed AI subscription on cost only above a specific usage threshold. Below that threshold, a $21 to $75 per-seat platform is cheaper once you count GPU idle time, MLOps staffing, and the breach risk from AI nobody is tracking. The sticker price on either path is the opening line, not the total.

What are the two paths enterprises are actually choosing between?

Most enterprise AI decisions boil down to two options. Subscribe to a closed platform like ChatGPT Enterprise or Microsoft 365 Copilot and let someone else run the infrastructure. Or self-host an open-weight model like Llama, Qwen, or GLM on GPUs you own or rent, and run the infrastructure yourself.

This gets pitched as a philosophical choice. It isn't. It's a total cost of ownership question, and the answer changes depending on how much you actually use the thing, how sensitive your data is, and whether you already have engineers who can run a GPU cluster without burning a quarter learning how.

What does a closed AI subscription actually cost per seat?

The headline numbers look simple. Microsoft's enterprise Copilot add-on lists at $30 per user per month. ChatGPT Enterprise has no public price card, but 2026 procurement reports converge on $45 to $75 per seat per month, roughly $60 on average.

Neither number is the real number.

Microsoft 365 Copilot requires a qualifying Microsoft 365 license underneath it, usually E3 or E5. Add that in, and account for Microsoft's July 2026 price increase on the base suite, and the true all-in cost lands at $69 to $90 per seat per month, not $30. ChatGPT Enterprise reportedly carries a roughly 150-seat minimum on an annual commitment, which puts the realistic entry point near $108,000 a year before a single custom workflow gets built. Compare that to ChatGPT Business at $20 per user per month with no seat minimum, and you can see why so many mid-size teams start there instead.

Multiply any of these across a real headcount and the gap compounds fast. A 500-person org on the low end of Copilot's all-in cost is paying over $400,000 a year. At the high end of ChatGPT Enterprise pricing, it's north of $450,000. That's before anyone touches the seat pricing stacked on top of it from every other SaaS tool in the building, the same per-seat math that shows up when you compare internal AI workflow costs against SaaS subscription costs.

What does self-hosting an open-weight model actually cost?

The self-hosting pitch is that open weights cost pennies per token, so running your own infrastructure should always win. The hardware alone makes the case look easy: a consumer-grade RTX 5090 runs about $2,000, a single A100 runs around $15,000, and a full cluster for a mid-sized model runs $30,000 to $240,000 depending on scale. An 8-GPU H100 reference system from Lenovo lists at $833,806.

That's the CapEx line. It is not the TCO line, and the gap between the two is where the real math nobody puts on the slide tends to live, as the self-hosting crossover math shows in more detail.

Three operating costs get left off the napkin math most often:

• Idle GPU time. Traffic variance means 20 to 40 percent of GPU capacity sits unused on a typical self-hosted deployment, which is paid-for compute doing nothing.
• DevOps and incident response. A self-hosted cluster needs 0.5 to 1.0 FTE dedicated to keeping it running and responding when it breaks.
• Eval, monitoring, and tuning. Add another 10 to 15 percent of GPU cost for ongoing evaluation and monitoring, plus $2,000 to $10,000 for the initial quantization and tuning work to get a model production-ready.

The staffing line is the one that quietly dominates the whole model. Senior AI and MLOps engineers in North America command $180,000 to $280,000 in base salary, with total compensation running $250,000 to $450,000, averaging around $340,000. A cluster needing even half an FTE at that comp level adds six figures a year to a bill that was supposed to be "basically free compute."

At realistic throughput, a Llama-8B deployment on a single $2-per-hour H100 lands around $18.70 per million tokens, roughly six times more expensive than Claude 3.5 Sonnet on input-heavy workloads, though competitive on output-heavy workloads, before operations costs get added on top. Self-hosting is not automatically cheap. It's cheap at scale, under the right conditions.

When does open-weight self-hosting actually win on cost?

A Carnegie Mellon study modeling 54 deployment scenarios across nine open-source models and six commercial APIs found break-even timelines that vary enormously by model size. Small models under 30 billion parameters, deployable on a single ~$2,000 consumer GPU, break even against premium commercial APIs in as little as 0.3 months. Medium models in the 70B to 120B range need 3.8 to 34 months on $15,000 to $30,000 hardware. Large models over 200B parameters can take up to 69.3 months, more than five and a half years, comparing Kimi-K2 against GPT-5.

The pattern is simple: break-even scales with model size and usage volume, not with ideology. The math only pencils once you cross a real, calculable break-even, as the real math on renting tokens versus owning hardware lays out in detail.

Hardware economics tell the same story from a different angle. Buying an 8-GPU H100 system at $833,806 only beats renting the equivalent capacity, at roughly $2.79 per GPU-hour, about $22.32 an hour for the full system, after roughly 38,900 hours of use, which is about 21 hours a day for five straight years. That break-even is calculated before staffing, networking, and storage costs get added. For anything short of sustained, near-constant utilization, renting beats owning.

What hidden costs don't closed platforms put in the brochure?

Closed platforms sell predictability. What they don't advertise is how expensive leaving gets. One enterprise reported $7.2 million in switching costs to migrate off a locked-in AI platform, not counting lost productivity during the transition, and 47 percent of enterprises say they're functionally trapped in their AI vendor relationship within three years. True exit costs tend to run 40 to 60 percent higher than enterprises initially project, once data migration, model retraining, integration rebuilds, and staff retraining are counted. That's what lock-in costs when the bill comes due, and it rarely shows up in the original procurement deck.

Price volatility compounds the problem. Microsoft's own July 2026 increase to the Microsoft 365 base suite, which Copilot pricing rides on top of, is a live example of a vendor resetting the floor after you've already built workflows on top of it.

What hidden costs does open-weight hosting leave out of the spec sheet?

Self-hosting has its own brochure problem, just pointed the other way. A 100-GPU on-premise cluster's total cost of ownership is dominated by power, cooling, networking, storage, and engineering talent, not the sticker price of the GPUs themselves. Server and networking costs can rival the cost of the compute nodes outright.

Idle time is the single largest financial drain in that model, and it's exactly why specialized cloud GPU rental becomes mathematically superior for bursty, unpredictable workloads instead of flat-out ownership. If your usage spikes around quarter-end or product launches and goes quiet otherwise, you're paying for a cluster that sits idle most of the year, the same idle-time tax described above, just restated at the hardware layer.

Why is shadow AI the cost nobody budgeted for either way?

This is the part both camps miss. Whether you subscribe to a closed platform or stand up your own cluster, employees are still going to paste company data into whatever free AI tool solves their immediate problem, and that behavior now has a measured price tag.

IBM's 2025 Cost of a Data Breach Report found that organizations with high levels of shadow AI paid $670,000 more per breach on average than those with low or no shadow AI, and 97 percent of AI-related breaches involved organizations that lacked proper AI access controls. One in five breached organizations had a breach linked specifically to shadow AI.

The governance gap behind those numbers is structural, not accidental:

• 63 percent of breached organizations have no AI governance policy, or one still in development.
• Of the organizations that do have a policy, only 34 percent audit regularly for unsanctioned AI use.
• 97 percent of AI-related breaches trace back to missing AI access controls.

None of that depends on which platform you picked for your official AI tool. It depends on whether anyone is watching what employees do on the side, a governance problem that sits on top of whichever cost model you choose.

How do seats, tokens, and sensitivity fit into a simple decision framework?

Instead of asking "open or closed," ask these four questions and let the usage data answer for you:

1. Token volume per month. Low, bursty usage favors a closed subscription or rented GPU capacity. Sustained, high-volume usage is what makes the self-hosting break-even curve bend in your favor.
2. Data residency and compliance needs. If regulatory or contractual requirements demand data never leave your infrastructure, that constraint can outweigh the cost math entirely.
3. Existing MLOps capacity. If you already have engineers who can run inference infrastructure, the marginal cost of adding a cluster is much lower than it looks on a cold start. If you'd be hiring from scratch at $250,000 to $450,000 in total comp, that hire is the real cost of entry.
4. Growth trajectory. A team scaling usage fast should model where it lands in 12 to 18 months, not where it sits today, because break-even is a function of cumulative volume, not a snapshot.

McKinsey's 2026 global AI survey found that about 20 percent of organizations already say AI-related operating costs, including token costs, are constraining their AI use, even as most plan to keep increasing investment. Cost uncertainty is a live operating constraint right now, not a planning exercise you solve once and forget.

Whichever direction a team takes, the gateway and governance layer sitting in front of the model matters as much as the model itself, which is where the self-hosted gateway comparison becomes relevant.

How do the real costs compare side by side?

For a mid-size organization, say 500 seats, the three-year picture looks roughly like this. Every number here is a band, not a quote.

• Closed platform path: 500 seats at $60 to $90 per seat per month, scaled over three years, lands somewhere between $1.08 million and $1.62 million in subscription cost alone, before switching-cost exposure of $7.2 million or more if the relationship goes sideways.
• Self-hosted path: An 8-GPU cluster at roughly $834,000 in hardware, plus three years of 0.5 to 1.0 FTE MLOps staffing at $250,000 to $450,000 a year, plus 20 to 40 percent idle-time overhead and 10 to 15 percent eval overhead on top of the GPU cost, lands in a comparable or higher range unless utilization stays consistently high across the full period.

Neither path is cheaper by default. The closed platform wins on predictability and low usage. Self-hosting wins once volume and utilization justify the fixed cost of hardware and headcount. Shadow AI adds a real, measured tail risk to both, regardless of which column you're counting from. The question worth asking isn't which brand to buy. It's where your actual usage sits on that curve, and whether anyone in the building is tracking the AI nobody approved.