We decided to open everything. We have 2 clusters running GLM-5.2 in production: one in the interior of Sao Paulo and another in Paraguay. Here is what runs, why, and the math that shows why self-hosting with open weights under the MIT license is insurance against the end of subsidized pricing.
The infrastructure that runs GLM-5.2
GLM-5.2 is a 744B MoE model with ~40B active parameters per token. In BF16, it needs ~1.49 TB of VRAM. In FP8, that drops to ~744 GB — which fits comfortably in 8x H200 (1,128 GB total). This is the official reference configuration.
Each rack: 8x H200 141 GB, NVLink 4 at 900 GB/s intra-node — mandatory for TP=8. In Paraguay, we run vLLM v0.23.0+ with FP8 KV cache and MTP of 5 draft tokens. In SP, SGLang v0.5.13.post1+ with chunked prefill 32768 and EAGLE 6 tokens. The engine choice depends on the workload: vLLM for aggregated throughput, SGLang for low single-stream latency.
Real throughput, not marketing numbers
In our tests, real throughput is 118 tok/s in single-stream and 1,215 tok/s aggregated with ~96 concurrent users. Median TTFT sits between 246-282 ms, TPOT between 7.16-7.54 ms. These are production numbers measured on a real cluster, not lab benchmarks with a 1-token prompt. These numbers come with FP8 KV cache enabled and MTP active — without those optimizations, throughput drops by half. The gap between 118 tok/s and 1,215 tok/s aggregated shows the batching gain: with ~96 concurrent users, the engine groups requests and extracts ~10x more throughput from the same hardware.
Why Paraguay and why interior SP
Paraguay: Itaipu generates 14 GW, 72.9 TWh in 2025, country 100% hydroelectric since December 2021. Decree 5306 of January 2026 fixes data center tariff at $0.028/kWh at 500 kV, with 15 years of preferential tariff. Law 7548/2025 grants 20 years of tax incentives. Asuncion averages 23 degrees, free cooling from May to September. Tigo fiber to Brazil, Asuncion-Sao Paulo 20-40 ms.
Interior SP: sub-5 ms to IX.br, Anhanguera-Bandeirantes fiber, land 40-70% cheaper. CPFL costs ~$0.14/kWh — 5x more expensive than Paraguay, but latency compensates for instant response.
The energy that matters
The node draws ~8 kW at the rack, ~10.4 kW with PUE 1.3, ~91,000 kWh/year at 100% usage. Paraguay: ~$2,548/year. SP: ~$12,740/year. The energy difference alone — ~$10K/year per node — pays for the Paraguay cluster in a few years. And that is just energy. CAPEX doubles the advantage.
The break-even math
CAPEX in Paraguay: ~$400-500K per server (Law 7548 exempts taxes). In Brazil: ~$700-800K (II 18% + IPI + PIS/COFINS 11.75% + ICMS 18% SP) — nearly double. The difference is not marginal: it is the difference between break-even in 1.2 years and 2.5 years.
At 1,200 tok/s, 24/7/365, that is ~37.8 billion output tokens/year. At real API price ($4.40/M output, $1.40/M input, 3:1 ratio), the equivalent is ~$325,080/year. That is the opportunity cost of not having your own cluster.
Break-even Paraguay: ~1.2-1.5 years at 100% utilization, ~2.5-3 years at 50%. Brazil: ~2.2-2.5 years and ~4.3-5 years. Even in the pessimistic scenario (Brazil, 50% utilization), the cluster pays for itself in under 5 years.
The API price is subsidized — and that is the point
The API price of $4.40/M output tokens is subsidized. According to Z.AI financial data, the company lost 4.72 billion yuan in 2025 (~$650M) on revenue of only ~$99M. It is burning IPO capital ($558M in January 2026) and VC money from Alibaba and Tencent. The real cost to serve a token (GPU CAPEX, energy, staff, depreciation) is 3-5x higher than the charged price. Market estimates indicate that NVIDIA's margin on the H200 is 70-80% — monopoly pricing.
This is not sustainable. A company that loses 6.5x its annual revenue does not survive without infinite external capital. And infinite external capital does not exist.
What happens when the subsidy ends
When IPO/VC capital runs out, API and plan prices must rise to cover real cost. Those depending on third-party infrastructure will feel the squeeze: price increases, rate limits, vendor lock-in. Self-hosting with open weights under the MIT license is insurance against this future. Even if Z.AI or any other AI company triples prices, the model remains yours — including the ability to improve it and develop your own foundation models based on top-tier models like GLM-5.2 and other open-source.
If you do not have the scale or financial capacity for your own rack: to enjoy the benefits of GLM-5.2, we recommend the coding plan or API. It is the shortest path to test in your workflow while the price is still subsidized.
Conclusion
Who controls the infra controls the cost. We run GLM-5.2 on 2 clusters, measure real throughput, and do the math. The break-even exists — and it arrives faster than you think when you stop paying someone else's subsidy. At Tech86, we help companies build the right AI infrastructure, in the right place, with the right math.