B4L (buildfor.life) prepared this planning tool as consultant to the investor, who would buy and own any servers and any data center. You pick where the GPUs run (colocation at CMU ITSC, the investor's own data center, rented cloud GPUs, or bought API tokens), the fleet, and how its GPU time is sold: as a token API, as GPU rental by the hour, as dedicated hosting per server, or for internal use. Next to the GPUs it can add virtual private servers (VPS) and dedicated servers rented by the month, either from hall B of the investor's data center or from a rented rack. The page then works out the upfront cost, the money in and out each year, payback, ROI, NPV and IRR, and it sets every preset business model side by side.
The figures date from October 2026 and come from the study documents (hardware, option A, option B, Thai compute, Thai models, summary, servers to rent) and from model/rent_vs_buy.py and model/server_packages.py; numbers in brackets link to the sources. Our own estimates are marked est., and figures that only a vendor states are marked vendor claim. The page is not an offer, a quote or a forecast. Demand is the decisive unknown: the utilization figures are estimates, so read the cautious case before the others.
Low takes the cheapest cited price and the best cited resale, high the opposite.
VPS and whole servers rented by the month, with the packages, builds and prices of the server packages note. They are a separate product line next to the GPUs.
Every figure on this card is an estimate (est.); all the other figures are listed under Assumptions.
Each row is a preset, calculated at the current demand, price end, horizon and finance settings. Revenue and EBITDA are for a mature year; payback, NPV and IRR are after tax over the horizon. Click a row to load it.
The table shows a mature year (from year 3) at the current settings.
Year 0 is the upfront cost. Utilization ramps up over years 1 and 2 (see Assumptions), and on a 10-year horizon the hardware is bought again in year 5 while the old fleet is sold at the resale band. Cash is counted after corporate income tax (with the BOI exemption when it is on), without financing and without inflation.
This table lists every figure the page uses, with its kind and source. Changed values are kept in the link, so a scenario can be shared with "Copy link".
Roles. B4L prepared this tool as consultant; the investor would buy, own and run the hardware and any data center. Nothing here is an offer or a forecast.
Demand. No Thai model maker sells a paid, Thailand-hosted API today, so there is no market data for Thai token volumes or premiums (Thai models). The demand switch and the utilization inputs are estimates. An 8x B200 at the interactive benchmark makes about 30 billion tokens a day at full load; the cautious case is the one to plan on.
Throughput. Interactive figures count input plus output tokens at 50 tokens/s per user; offline ceilings flatter owning. No B300 or GB300 benchmark is on file, so both use the B200 figures (est.). Thai models (30B mixture-of-experts) would likely run faster than gpt-oss-120b (our reading); Typhoon throughput on these GPUs was not found.
Comparison with model/results.md. The card "Against renting and API tokens" uses the model's basis (server price, support, ITSC rack or facility, power at 25% to 85% of nameplate; no staff, network, insurance, software or resale at five years), so one 8x B200 at ITSC reproduces its S2 rows. The rest of the page is all in.
Not included. The page leaves out import duty and VAT on servers, the integrator's margin, internet transit (unpriced at ITSC), sales and marketing, land, the grid connection, certification and design fees, working capital, financing, inflation and price trends, the BOI cap on the tax exemption, the Typhoon license review and US export checks (Thai compute). Rents moved fast in 2026 (H200 +143% since January, a news report)[24].
Servers to rent. The VPS and dedicated-server lines follow the server packages note and model/server_packages_results.md. A host S sells 240 VPS like Shared 2 and a host D sells 24 like Dedicated 4, with its spare Ceph space as block storage; a dedicated server rents at its package price. Support and platform staff grow along a straight line between the note's two-rack and ten-rack fleets (est.), and every instance sold pays for a leased IPv4 address, as in the note's fleet examples. In the investor's data center the lines pay hall B's facility charge per kW of peak draw, which the model derives from the whole 2 MW building, so the designed IT load above counts the GPU hall only. The servers have no resale value and are bought again in year 5 on a 10-year horizon. Setup fees, add-ons such as backups and Windows licenses, and the IPv6-only discount are left out. When the BOI switch is on, its exemption also covers the server lines, which BOI would have to confirm, since cloud services have their own promotion category.
Solar. The share covered hour by hour comes from an average day (a half-sine from 06:00 to 18:00 against a flat load, with one battery cycle a day). It is a placeholder until the site study's hourly data are in.