Starting a Frontier Lab in 2026: Compute, Cost and Capital, with a Bangalore Playbook
You want to build a model of your own and the agents that use it. How much compute does that take to train, to experiment and to serve? Where do the GPUs come from, how much money do you need, and who will provide it? This long bonus article works through the whole bill with interactive calculators, then looks closely at what changes if the startup is in Bangalore: India's subsidised GPUs, rupee prices, government funds and a realistic strategy.
1. Three ambitions, three budgets
"Frontier model startup" covers very different companies. The first decision is which race you are in, because it sets the compute bill by orders of magnitude.
2. What frontier costs now
Capital for new labs reached sizes that would have been unthinkable a few years ago. Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, raised about $2 billion at seed in mid-2025. In March 2026, Yann LeCun's AMI Labs raised $1.03 billion at a $3.5 billion pre-money valuation to build world models; in April, Ineffable Intelligence, founded by former DeepMind reinforcement-learning lead David Silver, raised $1.1 billion at a $5.1 billion valuation. A funding tracker puts the median frontier-lab round in 2026 at about $1.07 billion โ ๏ธ.
Seed rounds for new labs
Reported rounds, billions of dollars.
At the other end, DeepSeek showed how efficient a strong model can be. Its V3 technical report lists 2.788 million H800 GPU-hours for the final model (2.664 million for pre-training, 119,000 for context extension and 5,000 for post-training), which it values at $5.576 million at an assumed $2 per GPU-hour. That number is the final run only: it excludes the experiments that came before it, the data, the staff and the cluster itself. Analysts such as Nathan Lambert at Interconnects stress that the real cost of a lab producing such a model is many times higher.
3. The training bill, line by line
A useful rule of thumb: training a dense transformer takes about 6 ร N ร D floating-point operations, where N is the number of parameters (for mixture-of-experts models, the active parameters per token) and D the number of training tokens. Divide by what a GPU actually delivers, its peak speed times the share you achieve (model FLOPs utilisation, MFU), and you get GPU-hours. Both DeepSeek-V3 and Llama 3 405B work out to roughly 35% MFU on this rule, which we use as the default.
Compute budget builder
Our model. Pick a preset or set your own. "Experiments" multiplies the final run to cover ablations, failed runs and smaller test models; "post-training" adds fine-tuning and RL as a share of pre-training.
Two things stand out. First, the final run is often the smaller part of the bill: labs typically spend several times its compute on experiments before committing to it. Second, chip choice and utilisation matter as much as price: moving from 25% to 45% MFU nearly halves the GPU-hours. That is the work of the engineers in kernels and compilers and training frameworks, and the Llama 3 story in Producing Intelligence.
4. Serving the model, and the agents
Training is a one-off; inference is forever. Once a model has users, serving it usually becomes the larger compute bill, and agents make it larger still. Gartner's March 2026 forecast put agentic workloads at 5 to 30 times the tokens per task of a standard chatbot exchange; a Stanford study of coding agents found them consuming about 1,000 times the tokens of code chat, with runs on the same task differing by up to 30 times.
Serving cost calculator
Our model: self-hosted serving on rented GPUs. Throughput per GPU is an assumption that varies hugely with model size, batch size and hardware.
The lesson for founders: price your product per task, not per token, and budget for agent token counts to vary widely. Related reading: the inference engineers and the collapse in token prices.
5. Where to get the GPUs
The four routes from Part 1 apply to a startup, but in a different order. Few new labs build or buy a cluster in year one; most rent, and the question is on what terms.
Startup sourcing options
Tap an option. Prices are 2026 tracker ranges from Part 6; terms vary by deal.
Cloud credit programmes, such as AWS Activate, Google for Startups Cloud, Microsoft for Startups and NVIDIA Inception, can cover early experiments, but amounts and eligibility change often and are small next to a pre-training run. Larger labs have taken investment from cloud and chip companies that is partly delivered as compute; Google's and Amazon's investments in Anthropic are the best-known examples. Such deals bring capacity but also tie the lab to one provider.
6. How much money, and from whom
Who funds a frontier lab
Typical sources and what they bring. Tap one.
Runway calculator
Our model. How long a round lasts given compute and team costs, and how much of the company it costs.
7. The Bangalore playbook
For a startup in Bangalore, four things change: access to subsidised national compute, rupee prices for commercial GPUs, government and domestic funding, and the market it should target. Bangalore also brings India's deepest pool of engineering talent at a fraction of US salaries.
7.1 The IndiaAI Mission's compute
The IndiaAI Mission, approved in 2024 with a โน10,372 crore outlay over five years, had onboarded 38,231 GPUs by March 2026 through a common compute portal, rented to startups, researchers and academic institutions at subsidised rates reported at about โน65โ67 per GPU-hour. Reports in 2026 said startups building foundation models would receive a 100% compute subsidy, against 40% for other AI workloads โ ๏ธ (anonymous sources).
What an H100-hour costs a Bangalore startup
Rupees per GPU-hour. US prices converted at โน88/$ (assumed).
7.2 The selected model builders
The government selected startups to build foundation models in phases: Sarvam AI, Soket AI Labs, Gnani.ai and Gan.ai first, and a later report counted 20 proposals, 12 large and 8 small language models. Sarvam is furthest along: it was allocated 4,096 H100 GPUs for six months, with a subsidy of โน98.68 crore on a compute bill of โน246.71 crore. Analytics India Magazine reported in October 2026 that the other selected startups had not yet received their GPUs โ ๏ธ, a reminder that allocation and delivery are different things.
India's AI compute and model builders
A schematic map; tap a dot.
7.3 Funding from India
Beyond compute, Indian founders can draw on domestic venture capital, global funds with Indian teams, corporate backers and new government vehicles. The โน1 lakh crore Research, Development and Innovation (RDI) scheme, approved in July 2025, lends long-term at low or zero interest through second-level fund managers; its first cohort in September 2026 backed 22 projects involving โน4,744 crore of R&D, 14 of them startups. The IVCA's Bharat DeepTech Report counts $11.4 billion of cumulative deep-tech funding from 2015 to mid-2026, with $2.96 billion across 189 deals in 2025 alone. That is a deep pool for a specialised lab, and a shallow one next to a single US frontier round.
7.4 Chips, power and policy
- Chip access: the US AI diffusion rule, which would have placed India in a middle tier with chip quotas, was rescinded in May 2025; as of mid-2026 no replacement had been published. Indian buyers can import advanced NVIDIA systems, subject to normal export licensing.
- Data centers: Karnataka approved a Sustainable Data Centre Policy for 2026โ2031 targeting 1 GW of IT load and a 10โ12% share of India's AI-ready data center capacity, with capital subsidies, power-tariff benefits, electricity duty exemptions and a one-time โน20 crore payment for very efficient sites, and a push beyond Bengaluru to places such as Mangaluru.
- Talent: Bangalore's engineering base lowers team costs sharply compared with San Francisco, though experienced large-scale training engineers remain scarce everywhere.
7.5 A realistic strategy
8. Three founder plans
From seed to serving in 24 months
Illustrative plans built with this article's calculators. Not predictions.
9. Risks
10. Founder's checklist
11. Reading the numbers
- Training anchors
- DeepSeek-V3: 2.788M H800 GPU-hours, $5.576M at an assumed $2/hr (technical report); 37B active parameters, 14.8T tokens. Llama 3 405B: ~30.8M GPU-hours, 15.6T tokens (Meta). The ~35% MFU default is our back-calculation from these.
- Funding
- Thinking Machines ~$2B seed (mid-2025); AMI Labs $1.03B at $3.5B pre-money (Mar 2026); Ineffable Intelligence $1.1B at $5.1B (Apr 2026); median 2026 frontier round ~$1.07B (tracker โ ๏ธ).
- Agents
- 5โ30ร tokens per task vs chat (Gartner, Mar 2026); ~1,000ร for coding agents vs code chat (Stanford); these measure different things.
- India
- โน10,372 crore IndiaAI Mission; 38,231 GPUs, 190 projects (Mar 2026); ~โน65โ67/GPU-hr subsidised; 100% foundation-model subsidy reported โ ๏ธ; Sarvam 4,096 H100s, โน98.68 crore of โน246.71 crore; others' GPUs delayed per AIM โ ๏ธ; Yotta H100 โน263โ386/GPU-hr list; RDI scheme โน1 lakh crore, first cohort 22 projects; $11.4B deep-tech funding to mid-2026 (IVCA); diffusion rule rescinded May 2025; Karnataka data center policy 2026โ31.
- Models
- All calculators are simplified; they ignore data costs, failures, financing and taxes unless stated.
- Training and agents
- DeepSeek-V3 technical report
- Interconnects: the actual cost of training frontier models
- Stanford: how do AI agents spend your money?
- Crusoe: tokenomics in the age of agentic inference
- Funding
- Futurum: AMI raises $1B seed
- Ineffable Intelligence raises $1.1B seed
- Frontier AI labs funding trends (tracker)
- India
- PIB: IndiaAI Mission compute
- AIM: IndiaAI, where are the GPUs?
- Inc42: 100% compute subsidy for foundation models
- Kotak Neo: 20 indigenous models
- Yotta Shakti Cloud: bare-metal pricing
- Business Today: RDI fund backs 22 projects
- YourStory: Indian deep-tech investment $11.4B
- Takshashila: AI diffusion framework repealed
- The News Minute: Karnataka data centre policy
- On this site
- Founding an AGI Startup series ยท Part 6: the economics ยท India AI