AugmaThe AI players
A handful of companies build it. Everyone else decides how to use it.
Who makes the AI that large companies run on, how the market is split between them, what a few big organisations have actually done with it, and what that means for a business our size. Every figure links to the page it came from.
The landscape
Four layers, and most companies buy through the middle.
Model makers
Closed frontier models, sold as an assistant for staff and as an API for builders.
OpenAI
ChatGPT for business teams and the GPT models by API; also offered through Microsoft and, for agents, through Amazon Bedrock.
AWS, Amazon Bedrock ↗Anthropic
Claude for Enterprise: large context, single sign-on, role permissions, audit logs, and a GitHub integration for engineering teams.
Claude for Enterprise, Sept 2024 ↗Clouds and platforms
Where most companies actually buy: many models behind one bill, one security model and one set of agent tools.
Microsoft
Copilot inside Microsoft 365, and Foundry: over 11,000 models from OpenAI, Anthropic, Meta, xAI and others, plus an agent service.
Microsoft Foundry ↗Makes the Gemini models and runs them: Gemini Enterprise for staff across Workspace, and Vertex AI with Google, open and third-party models.
Google Cloud, Vertex AI ↗Amazon
Bedrock: a choice of foundation models from many companies, agents, knowledge bases and fine-tuning, with compliance such as HIPAA eligibility.
AWS, Amazon Bedrock ↗Open models
Weights you can download, run on your own servers and adapt to your own language and data.
Meta · Llama
Llama 4 Scout and Maverick, released for download in April 2025.
Meta AI, Llama 4, Apr 2025 ↗Mistral
A European maker that leads with sovereignty: self-hosted on-premises, its own EU cloud, or through partner clouds.
Mistral AI ↗DeepSeek
DeepSeek-R1, a reasoning model released under the MIT licence in January 2025.
DeepSeek, R1 release, Jan 2025 ↗Alibaba · Qwen
Alibaba says it has open-sourced more than 460 Qwen models, with over 300,000 derivatives built on them.
Fortune, Aug 2026 ↗Infrastructure
The chips and systems underneath every one of the above.
Nvidia
Data-centre revenue of $41.1 billion in one quarter (to July 2025), and partnerships on sovereign AI systems with several European countries.
Nvidia, Q2 FY2026 results, Aug 2025 ↗Listed in no order of merit. The lines blur: Google both makes models and runs a cloud, Microsoft and Amazon resell other makers' models, and several closed makers also publish open ones.
Where the money goes
Spending tripled in a year, and the lead keeps changing hands.
- $37Bspent by enterprises on generative AI in 2025, up 3.2 times from $11.5 billion in 2024Menlo Ventures, State of Generative AI in the Enterprise, Dec 2025 ↗
- 89%of enterprise LLM API usage runs on closed models; open models hold 11%Menlo Ventures, State of Generative AI in the Enterprise, Dec 2025 ↗
- 37%of 100 surveyed CIOs run five or more models in production, up from 29%a16z, enterprise CIO survey, Jun 2025 ↗
Share of enterprise LLM API usage, end of 2025
In 2023 the same survey put OpenAI at 50% and Google at 7%. One survey, one moment: treat it as a direction, not a verdict. Menlo Ventures, State of Generative AI in the Enterprise, Dec 2025 ↗
How big companies use it
Five organisations, five different first moves.
- Payments · Sweden
Klarna
Put an AI assistant on the front line of customer service chat, in 35+ languages across 23 markets.
2.3M conversations in the first month, two-thirds of all chats; resolution time fell from 11 minutes to under 2, with satisfaction on par with human agents.Klarna press release, Feb 2024 ↗ - Banking · USA
JPMorganChase
Built LLM Suite, an internal portal that puts large language models behind the bank's own controls, for contracts, presentations, client emails and reports.
200,000+ employees use it, across legal, sales and client services.JPMorganChase via PR Newswire, Jun 2025 ↗ - Retail · USA
Walmart
Used several large language models to create and clean the product catalogue that search, inventory and delivery all depend on.
850M catalogue data points created or improved; the CEO said doing it by hand would have needed nearly 100 times the headcount.Diginomica, Aug 2024 ↗ - Commerce software · Canada
Shopify
Changed the culture rather than one process: the CEO wrote that reflexive AI use is now a baseline expectation for everyone.
Policy: teams must show why AI can't do a job before asking for more people, and AI use enters performance reviews.Digital Commerce 360, Apr 2025 ↗ - Banking · Thailand
KBTG (Kasikornbank)
Built THaLLE, a finance model fine-tuned from an open Qwen model on CFA-style questions, and released it openly under Apache 2.0.
71.7% on a CFA benchmark, ahead of its open base model (68.3%) but behind GPT-4o (87.9%), as its own report states.KBTG Labs on Hugging Face, Jun 2024 ↗
These are each company's own stated results, not independent audits. Notice how different the moves are: a customer-facing assistant, an internal portal, a data clean-up, a cultural rule, a home-grown open model.
Where it's heading
Five shifts worth watching.
- AgentsEvery platform above now sells tools to build agents that act, not just answer: Foundry Agent Service, Bedrock agents, Google's agent platform, Mistral Studio.Microsoft Foundry ↗Amazon Bedrock ↗
- Open vs closedOpen models closed the benchmark gap to about 1.7% in a year, yet enterprises still run 89% of usage on closed ones. Quality caught up; convenience and support have not.Stanford HAI, AI Index 2025 ↗Menlo Ventures, State of Generative AI in the Enterprise, Dec 2025 ↗
- Many models, not oneCIOs pick models per task, and the reason they give is fit, not fear of lock-in. The clouds are built for this: one account, many makers.a16z, enterprise CIO survey, Jun 2025 ↗
- Sovereign and local AICountries and companies want models that speak their language and data that stays home. Europe is building national systems with Nvidia; in Thailand, SCB 10X's Typhoon offers open Thai models, including Isan speech.Nvidia, Aug 2025 ↗Typhoon, SCB 10X ↗
- Cost per token fallingRunning a model at GPT-3.5's level got over 280 times cheaper between November 2022 and October 2024. What was too costly to try last year may be cheap now.Stanford HAI, AI Index 2025 ↗
What this means for you
You don't need a bank's budget. You need one clear job.
For a small or mid-sized business in Thailand or Southeast Asia, the honest picture is good news with a caveat. The good news: the same models the banks use are on sale by the token, prices keep falling, and there are open models that read Thai well and can run on your own machine when the data must stay in-house.
The caveat: none of the results above came from picking the 'best' model. They came from choosing one job, preparing the data behind it and measuring the change. Don't marry one vendor; most large companies don't. Start with what your team already uses, keep a person in the loop, and switch models when the work tells you to.
Not sure which of these fits your work?
We're builders who work with several of these models every day, and we have no vendor to push. Tell us what you do now, and we'll think it through with you.