Own your inference.
An 8-GPU NVIDIA B200 server, co-owned pro rata by a group of organizations, running GLM-5.2. You use it to replace ~90% of your API bill and pay much less for your inference over 10 years and beyond.
For organizations spending $10–20K+/month on AI APIs · minimum stake of one-fifth · hosted and operated in Canada.
Five co-owner organizations
An 8-GPU B200 server, co-owned pro rata by 5 organizations — $1.0M turnkey, or $200,000 per seat. The server is worth 37 sellable PTU (provisioned throughput units), and each seat gets a whole number: 7 PTU at five co-owners, 4 PTU at nine; the remaining PTU are resold for the co-ownership. What that gives in tokens per second depends on the input/output mix — on pure output, one seat in five holds about 1,320 tok/s and one seat in nine 760 tok/s; on an agentic workload (10 tokens read per token written, 85% cached), about 670 and 380 tok/s. An autonomous AI coding agent (e.g., Claude Code) generating code continuously runs at roughly 50 tokens/second — that's 13 concurrent AI agents per seat at five co-owners and 7 at nine, on an agentic workload.
The calculation is the one behind the syndicated plan: the server is worth 37 sellable PTU (about 41 in total, with 10% held in reserve for peaks), and each seat gets a whole number of them. What changes is the heaviest line item. In syndication, the plan carries the GPU lease — $36,000 a month. In co-ownership the GPUs are paid for once, and only operations remain: colocation, systems administration, network, observability, security and failover, or $7,600 a month for the whole server, which is about $205 per PTU, versus $1,126 per PTU of server cost on the syndicated plan.
That amount does not depend on the volume consumed or on the token mix. The output throughput it buys does: input tokens, cached or not, also consume PTU capacity, and the more context your requests re-read, the fewer output tokens per second remain for the same number of PTU.
| Workload | Output throughput, one seat in five | Output throughput, one seat in nine | Output equivalent |
|---|---|---|---|
| Output only | 1,323 tok/s | 756 tok/s | $0.41 CAD/M |
| 10:1, 85% cached | 669 tok/s | 382 tok/s | $0.82 CAD/M |
| 30:1, 85% cached | 336 tok/s | 192 tok/s | $1.63 CAD/M |
| 20:1, 30% cached | 228 tok/s | 130 tok/s | $2.40 CAD/M |
One PTU produces about 189 output tokens per second, or processes 595 uncached input tokens per second, or 3,205 cached input tokens; a request consumes a mix of the three. “10:1” means 10 tokens read per token written. The output equivalent is the operating cost per PTU, $205 a month, divided by the tokens a PTU generates running 24/7, with input counted in the workload; the PTU not allocated to seats, 2 at five co-owners and 1 at nine, are resold for the co-ownership. There is no context ceiling to choose: requests are accepted up to 1M tokens, and traffic above your PTU goes through on a best-effort basis, without a throughput guarantee.
Compare in detail
Not all tokens are equal. Input (your context) is read in parallel — the hardware processes hundreds of thousands of tokens per second. Output is generated one token at a time, and each token requires a full pass through the model. Producing a token genuinely costs several times more than reading one — which is why usage-based providers price output 3–5× higher than input.
In agentic workloads, input dominates: every call re-reads the full history — several thousand tokens read for every thousand generated. Its real marginal cost on a B200 server: roughly $0.09 CAD per million tokens. Anthropic bills input at 3 $ US/M (~$4.11 CAD) and Z.ai at 1.40 $ US/M (~$1.92 CAD) — that's 20 to 45× the actual cost. Billed by usage, that re-reading is the provider's most profitable line item.
| Anthropic (Sonnet)* | Z.ai (GLM 5.2)* | Alibaba Cloud (GLM 5.2)* | Your server (GLM 5.2) | |
|---|---|---|---|---|
| Input | $3 US/M (~$4.11 CAD) | $1.40 US/M (~$1.92 CAD) | $0.97 US/M (~$1.32 CAD) | counted in your PTU† |
| Output | $15 US/M (~$20.55 CAD) | $4.40 US/M (~$6.03 CAD) | $3.04 US/M (~$4.16 CAD) | ~12 months: ~$7 CAD/M, then ~$0.41 CAD/M‡ |
| Long context | up to 1M tokens | up to 1M tokens | up to 1M tokens | up to 1M tokens per request§ |
* September 2026 prices, at regular price with no promotion.
† Input is not billed per token, but it is not free either: input tokens, cached or not, consume your PTU capacity, and the number of PTU is sized on your peak. One PTU processes about 595 uncached input tokens per second, or 3,205 cached — an agentic loop re-reads its whole history on every turn, and the prefix cache makes that re-read about five times lighter. The monthly amount stays fixed, whatever volume is actually consumed.
‡ Funds capital + 10% repayment for the first ~12 months; becomes an ongoing cost at full utilization afterward — $0.41/M corresponds to output only, $0.82/M to an agentic workload of 10 tokens read per token written, 85% cached.
§ No context is reserved in advance: a long request simply counts more input tokens, hence more PTU. Requests are accepted up to 1M tokens; traffic above your PTU goes through on a best-effort basis, without a throughput guarantee.
GLM 5.2 supports 1M tokens of context natively, and its sparse attention keeps processing cost from climbing steeply as context grows. Your seat holds PTU, sized on your peak with input counted alongside output: an AI agent re-reads its entire history on every call, and that re-reading is exactly what a usage-billed provider bills in full, every turn. On your server it occupies PTU capacity, heavily lightened by the cache, without changing the month's amount. Requests are accepted up to 1M tokens; beyond your PTU, they go through on a best-effort basis.
Inference you own, with none of the logistics
GLM-5.2 (744B MoE, FP8) runs on the rig via vLLM + LiteLLM, for your agentic clients (Claude Code / OpenCode). About 90% of your tokens shift to inference you own — equivalent to roughly $7/M at retail for the first ~12 months (while capital + 10% is repaid), then ~$0.41/M in ongoing costs at full utilization, on output only.
Your monthly plan has two amounts: server operations, billed at cost with no margin, and the Agent Libre fee, from $483 to $312 per PTU per month, which drops as your number of PTU grows. Since you own the GPUs, GPU rental, which is 86% of operations, disappears from your bill. See the pricing grid
Open-weight models you own, routing you control, and data that stays under the members' control. The commitment is to the intelligence layer, not the silicon — renewed with every performance improvement.
The server is installed in colocation in Canada, operated 24/7 by Agent Libre and its partners — power, cooling, network, and maintenance included. No logistics on your end.
The surplus is resold as tokens on marketplaces such as OpenRouter, for the co-ownership's benefit: $156,000 in year one, $640,000 over 10 years.
Prompt analysis and optimization, protection against sensitive-data leaks and token theft, no model training.
The purchase decision rests on real measurements.
We don't commit capital based on a spreadsheet. A one-week pilot, on your real workload, measures how much of your API bill GLM-5.2 actually replaces — before any purchase.
We measure the real substitution rate against your bill, the $/M tokens per organization, and throughput across co-owners. Marginal cost near zero.
Capital is committed only if the pilot confirms the threshold: real substitution rate, measured $/M tokens per organization, and negotiated hardware price.
No capital at risk until the pilot confirms your real savings. Once the measurements are in hand, the decision becomes mechanical.
The hardware you own
Reference configuration — Supermicro H14 10U server, air-cooled NVIDIA HGX B200 platform.
Measure your savings before committing any capital.
Book 30 minutes: we'll set up the one-week pilot, estimate the substitution rate against your current bill, and walk you through the full spreadsheet.
Your company has two supply chains
The first one you already know. Materials come in, they are transported, stored, transformed, delivered. Every step has a supplier, a lead time, a unit cost. When the margin tightens, you walk down that chain step by step, and you generally know where to look.
The second one moves information instead of materials. A price request comes in and becomes a quote. An order comes in and becomes a production order, an invoice, a confirmed delivery. A site report becomes a change order, or it does not. This chain runs through every one of your departments and it employs a large share of your people.
Let’s talk about your AI bill.
Describe your AI agents, your orchestrator and your current bill.