Hardware: what to buy, what it serves, what it costs against renting

Research note, October 27, 2026, by Alexander Grabowski (B4L). Sources were retrieved on October 27, 2026. Figures are cited, and our own estimates are marked.

This note covers the GPU servers the investor could buy, the models each can serve and what owning them costs against renting. B4L prepared it in its consulting role for the investor and does not buy or operate hardware itself. Money is converted at 33 THB per USD, the B4L planning rate. Because the sources quote server prices in USD, this note gives USD first.

In short

• An 8-GPU server costs USD 164,000 to 197,000 (5.4 to 6.5 million THB) with RTX PRO 6000 GPUs and USD 360,000 to 565,000 (11.9 to 18.6 million THB) with B200s, at US reseller prices in October 2026[20][21][22].
• Maximum power is 6.8 kW for the RTX PRO 6000 server, 10.2 kW for the H200 and 14.3 kW for the B200, all air-cooled[20][188][189]. A GB300 NVL72 rack draws 135 kW and needs liquid cooling[190].
• At the time-of-use rates of the Provincial Electricity Authority (PEA) from September 2026, a constant 24/7 load pays about 3.81 THB/kWh, including the demand charge, Ft and VAT (our arithmetic)[25][175].
• A B200 server owned by the investor at CMU ITSC beats renting the same GPUs on demand at Lambda (USD 6.69 per GPU-hour) above 23 to 43% utilization over five years (our model)[23].
• API tokens are the hardest baseline. OpenRouter sells gpt-oss-120b at USD 0.03 per million input tokens and USD 0.15 per million output tokens[19]. A B200 server beats that only above 23 to 41% interactive load (7 to 13 billion tokens a day), and RTX PRO 6000 and H200 servers need over 80% at best (our model).
• Rents for the H200 have risen 143% and for the B200 111% since January 1, 2026, according to Ornn Data as cited by a16z (a news report)[24], and some B300 builds quote lead times of 30 weeks[20].
• We recommend that the investor start with one 8x RTX PRO 6000 server for Thai models and private data, keep using APIs for general models, and design its data center for B300 and Rubin.

Server options

NVIDIA says DGX B300 systems are shipping now[194]. On January 5, 2026, it said Rubin was in full production, with partner products from the second half of 2026[204]. At about 24 kW per box, Rubin NVL8 needs 1.7 times the power of a DGX B200 (our arithmetic), so a hall built from 2027 on should plan for 25 kW and more per server and for racks ready for liquid cooling (our estimate; see option B).

AMD's MI355X (288 GB per GPU, 1,400 W board power)[205] sells as an 8-GPU platform for USD 300,000 to 390,000[206]. We do not model it, because the Thai models and the benchmarks below center on NVIDIA. The desk systems are development tools rather than shared servers.

Prices and supply

We found no NVIDIA list prices and no Thai distributor's price list; all the server prices come from resellers, an aggregator or Lenovo. For one 8x B200 they run from USD 360,000 to 470,000 (Rack Report)[21] and USD 499,000 with 3-year on-site support (a published reseller quote)[207] up to USD 550,475 (Lenovo, June 15, 2026)[196]. Mercatus calls allocation, rather than price, the defining feature of B200 buying in 2026 and puts a DGX B200 at about USD 515,000 list[208]. There is one Thai data point: Kaohoon International reported in February 2025 that Siam AI paid about 11 million THB (USD 333,000) per DGX B200[141].

Prices are volatile. NVIDIA raised the price of the 128 GB DGX Spark nearly 75% in a year, citing memory prices[200]. A used HGX H200 was valued at about USD 348,500 in August 2026, above the new price of roughly USD 340,000, which the Rack Report reads as a sign of scarce new supply[21]. In October 2026 lead times were 3 to 5 weeks for RTX PRO 6000 servers and 6 to 10 weeks for H200; the B300 was out of stock on the Supermicro eStore, at 8 to 12 weeks from buffer stock or 30 weeks and more when configured to order[20].

Thai retail listings are unusable as prices. ShoppingPC in Bangkok lists an 8x H200 Supermicro SYS-821GE-TNHR at 30 million THB and an 8x B300 Supermicro at 120 million THB[209]. The H200 listing comes to USD 909,000, about 3.2 times the US B2B price of USD 284,770 for the same model[20], so we read these listings as anchors for a quote. The workable route is a written quote from a Thai partner of Supermicro, Dell or Lenovo (Lenovo Thailand helped launch Mahidol University's H200 system in January 2026[210]), checked against the US price plus freight, duty, VAT and margin. The note on Thai compute covers US export rules.

What fits on what

The Thai models the investor would serve first are small. Typhoon 2.5 and the national ThaiLLM-30B base share the Qwen3-30B-A3B architecture, and the note on Thai models gives the full list with licenses. They fit on one GPU. Memory decides only for models of 671B to 1T parameters, which need an NVLink HGX node; Thai AI Cloud says a 671B model loads on one HGX B300 with room for a long KV cache (vendor claim)[130].

Throughput

Offline ceilings queue all requests at once and report the maximum output per GPU[132]. Interactive figures hold each user at 50 tokens/s in a chat with 8k input and 1k output tokens, and they count input plus output. The two kinds of figure must not be mixed[213].

MLPerf Inference v6.1 (offline) measured 91,487 tokens/s on gpt-oss-120B for a whole 8x B200 system[217]. At the interactive point, one 8x B200 server handles about 30.8 billion tokens a day, roughly 790 users at once (our arithmetic: 8 x 44,489 x 86,400; 4,943 output tokens/s per GPU / 50).

One RTX PRO 6000 serving gpt-oss-120b to 50 concurrent prompts delivered about 1,525 output tokens/s, roughly 30 tokens/s per user[131]. Its GPUs talk over PCIe instead of NVLink, and a field report on 8x RTX PRO 6000 (Max-Q variant) found prefill bound by communication at about 7,200 tokens/s, with the GPUs at 220 W of 600 W[218]. Qwen3-30B-A3B reaches 9,938 output tokens/s on one card but 3,985 per GPU when spread across eight[132]. In our reading, the box is better suited to many copies of 30B to 120B models than to one model split across eight cards.

For Typhoon we found only SCB 10X's claim that one H100 processes over 3,000 tokens/s at 64 concurrent requests[219], and no independent benchmark on RTX PRO 6000 or B200.

Buy versus rent

The comparison comes from model/rent_vs_buy.py, with results in model/results.md in the repository. It compares one server owned by the investor, at ITSC or in the investor's own data center, with renting the same eight GPUs and with buying the same tokens from an API, over five years and over three. The inputs are these:

• Prices are as above. Support ranges from Supermicro's 5-year warranty extension (USD 151,104, about 30% of a B300 server)[20] up to Lenovo's 12% of system cost per year[196].
• Resale value after three years comes from the Rack Report's bands for 2029 (B200 16 to 46%, H200 7 to 37%)[21]; there is none after five years, as in Lenovo's model.
• ITSC colocation costs 20,000 THB a month for a 42U rack (2021 tariff)[14][220]. The investor's own data center is a zero placeholder here and is priced in option B.
• A server draws 25% of nameplate power at idle and 85% at full load. PUE (total facility power over IT power) is 1.6 at ITSC and 1.2 in a modern data center built by the investor (our estimates).

PEA's type 4.2 time-of-use tariff at 22 to 33 kV charges, from September 2026 bills and before Ft and VAT, 4.1839 THB/kWh in the peak window (Monday to Friday, 09:00 to 22:00), 2.6037 THB/kWh off-peak and 132.93 THB per kW of monthly peak demand[25]. Ft for September to December 2026 is 0.1623 THB/kWh[175]. For a flat load, 0.387 x 4.1839 + 0.613 x 2.6037 = 3.215 for energy, plus 132.93 / 730 h = 0.182 for demand, plus 0.162 for Ft, gives 3.560 THB/kWh, or 3.81 with 7% VAT (our arithmetic, with a peak share of 65 of 168 hours and holidays ignored). A B200 server at full load at ITSC therefore uses about 170,000 kWh a year, which costs 650,000 THB (USD 19,700), and the rack adds 240,000 THB (our estimates).

The results for five years at ITSC, from the best case for the investor to the worst, are these:

Rents per GPU-hour were read on October 27, 2026, from CoreWeave[221], Lambda[23] and Bargo[222]. The Siam.AI H200 price comes from an undated page[61], and Runyard gives 74.69 THB an hour for the RTX PRO 6000[223].

Over three years, owning is harder (B200 against Lambda: 21 to 53%). The investor's own data center at PUE 1.2 barely helps (21 to 40%), because the server price and support dominate. Renting 24/7 is the most expensive option: Thai AI Cloud's dedicated 8x B300 node at USD 30,900 a month[130] comes to 61.2 million THB (USD 1.85 million) over five years, about three times the best case for owning a B200.

Lenovo's vendor analysis points the same way, and more strongly. At 100% utilization an 8x H200 breaks even against Azure on demand in about 5.2 months, and an 8x B200 beats AWS on demand (USD 114.27 an hour) above about 5.3 hours a day[196]. It is a server vendor's paper, however; its B200 conclusion names Google Cloud though the calculation uses the AWS rate, and hyperscalers charge USD 14.24 per B200-hour against USD 8.19 at neoclouds (CCIR index)[224].

The token baseline is the hardest. OpenRouter charges, per million input and output tokens, USD 0.03 and 0.15 for gpt-oss-120b[19] and 0.09 and 0.30 for Qwen3-30B-A3B[108]. SemiAnalysis puts gpt-oss-120b on an owned B200 at about USD 0.011 per million tokens at hyperscaler volume[213], so API prices sit close to the cost at hyperscale. In our model, a B200 owned by the investor at 50% interactive load costs 0.67 to 1.18 THB (USD 0.020 to 0.036) per million tokens, against 1.43 THB (USD 0.043) at the API; an RTX PRO 6000 server on gpt-oss-120b costs 8.2 to 11.2 THB against 4.4 THB. Saturated batch work flatters owning: at NVIDIA's offline rates, the RTX PRO 6000 server on Qwen3-30B-A3B beats the API above 5 to 8% utilization. What the API does not give is control. SCB 10X's Typhoon API is a free research showcase without formal support, and SCB 10X collects its usage data[126].

Reports on the rise in rents in 2026 differ. Ornn Data via a16z shows the H200 up 143%, the B200 111% and the RTX PRO 6000 73% since January 1[24], while Stocktwits reports 160, 120 and about 65%[225]. Bargo's average shows only 7% year on year for the B200[222]. We read these figures as a direction only.

Lifetime, reliability, support

Microsoft, Alphabet and Oracle depreciate servers over six years; where filings name GPUs, the median life is five years[226]. The Rack Report forecasts an HGX B200 at 16 to 46% of its new price in 2029 and 6 to 24% in 2030, and an HGX H200 at 7 to 37% and 4 to 9%[21]. We recommend that the investor plan a book life of five years and an economic life of three to four years, followed by batch and development use (our judgment).

The DGX B200 includes three years of Enterprise Business-Standard support[189]. Supermicro's base warranty is three years of parts and labor; three-year on-site next-business-day service adds USD 27,016 (about 16%) to the RTX PRO 6000 server[20].

Failures are routine. Meta's training of Llama 3 405B on up to 16,000 H100 GPUs had 466 job interruptions in 54 days; faulty GPUs caused 148 (30.1%) and HBM3 memory 72 (17.2%)[227]. An NCSA study of 1,056 A100 and H100 GPUs found H100 memory 3.2 times less resilient than A100 memory, and it estimates that 5% overprovisioning keeps job-level availability at 99.9%[228]. One server has no spare, so we recommend that the investor keep an API or cloud fallback and, from the second server on, about 5% spare capacity.

Dell saw 8x B200 GPUs near peak power throughout MLPerf inference[229], while one user reported an 8x B200 node at 6.5 to 7 kW at the wall (anecdotal)[230]. Hence our draw range of 25 to 85%, with circuits sized for nameplate.

Networking for a small cluster

NVLink carries 14.4 TB/s inside a DGX B200, which ships with eight 400 Gb/s ConnectX-7 ports[189]; RTX PRO 6000 servers have only PCIe. NVIDIA's enterprise reference architecture for HGX H100, H200 and B200 uses Spectrum-X Ethernet[231], so InfiniBand is not required. Splitting prefill and decode across GPUs needs RDMA; without it, time to first token is 200 to 500 times worse[232].

For one to four GPU servers, each holding whole copies of its models, a 25 or 100 Gb/s front-end network and a few hundred TB of NVMe storage are enough (our estimate). ITSC forbids private circuits into its data center and cables between racks without its approval[14], and CMU's port at the BKNIX Chiang Mai exchange is 10 Gbit/s[233]. The note on option A has more.

What this means for the investor

This section is our judgment as advisers; the decision is the investor's.

1. We recommend that the investor buy one 8x RTX PRO 6000 server first. It is the cheapest option (5.4 to 6.5 million THB), arrives in 3 to 5 weeks, fits ITSC more easily than a 14 kW HGX, and runs eight copies of models of the Typhoon 2.5 or ThaiLLM-30B class with data kept in Chiang Mai. It beats renting the same GPUs above 31 to 43% use but does not beat API tokens for general models, so the case for it rests on data control and on Thai models that no one sells through a commercial API.
2. The investor should keep using APIs for general open models, for peaks and for outages.
3. An HGX B200 or B300 belongs in a second stage, when demand passes the API break-even of 7 to 13 billion tokens a day, when a partner such as CMU commits to capacity, or when a 671B-class model must run in Thailand.
4. We advise against the H200. On gpt-oss-120b it gives about a fifth of the B200's interactive throughput (9,319 against 44,489 tokens/s per GPU)[214][213] at an overlapping price.
5. Option B should be designed for the next generation: DGX B300 at 30 kW per rack on air or 58 kW with rear-door heat exchangers[176], Rubin NVL8 at about 24 kW per box[203], and NVL72-class racks at 135 kW with liquid cooling[190].

Open questions

• Thai quotes. Get two quotes per server from Thai partners of Supermicro, Dell and Lenovo, each with the delivered price, lead time and allocation.
• Import duty and VAT on GPU servers. Ask the Thai Customs Department and the investor's accountant.
• BOI duty exemption. Could the project import servers duty-free in a BOI zone such as the Northern Science Park at CMU STeP[234]? Ask BOI's Chiang Mai office.
• On-site warranty in Chiang Mai. Ask the vendors about next-business-day service and local spares.
• ITSC terms. ITSC sets each colocation customer's power limit[14]. Ask its infrastructure section for that limit, whether power is included in the fee, whether the 2021 tariff still applies, and whether a rack can take 7 to 15 kW.
• Data-center tariff. A rate of 5 to 6 THB/kWh is under consideration, against the general rate of 3.95 THB[134]. At 6 THB/kWh a B200 server at ITSC would use about 1.0 million THB of power a year before VAT (our arithmetic). Ask PEA's northern region office and the Energy Regulatory Commission (ERC).
• Typhoon throughput on RTX PRO 6000 and B200 was not found; we recommend measuring it on rented GPUs before ordering.
• Rubin. Ask NVIDIA's Thai partners about prices and availability in Thailand.