United States: open weights, closed flagships

Bajío Ventures · October 2026 · Sovereign AI, 3 of 5

US labs have the most capable models in the world and do not publish them. What they do publish, since 2025, is a line of mid-sized models under permissive licences. The split is deliberate and, for industry, convenient: the range they give away is the one that fits on a server inside a plant.

What is available

Family	Latest open model	Size (parameters)	Licence	Where it runs
Google Gemma 4	2 Apr 2026	E2B, E4B, 26 billion and 31 billion	Apache 2.0	26 and 31 billion: one 80 GB H100 GPU unquantized; quantized, on consumer GPUs
OpenAI gpt-oss	5 Aug 2025	21 billion and 117 billion	Apache 2.0	The 117 billion model: one 80 GB GPU
IBM Granite 4.1	29 Apr 2026	3, 8 and 30 billion	Apache 2.0	No minimum published
Meta Muse Glimmer	Aug 2026	30 billion	Apache 2.0	24 to 32 GB of memory, quantized (InfoQ)
Nvidia Nemotron 3	Dec 2025 to Jun 2026	About 30, 120 and 550 billion	Nvidia’s own licence for the two smaller models; OpenMDW-1.1 for the largest	Optimized for its hardware
Microsoft Phi-4	Mar 2026	Up to 15 billion	MIT	Tested on A6000 GPUs and above

Three changes explain the table.

Google changed its licence. Earlier versions of Gemma had their own terms of use. Gemma 4 was released under Apache 2.0, and Google’s announcement uses the phrase “digital sovereignty” to explain it. The model is trained on more than 140 languages.

OpenAI published weights again. In August 2025 it released gpt-oss in two sizes, its first open language models since GPT-2. Since then it has published only gpt-oss-safeguard (October 2025), a variant fine-tuned for safety classification. We did not find a later general-purpose open model.

IBM bets on audits. Granite targets enterprise use. IBM presented the 4.0 generation as the first family of open models with ISO 42001 certification, the standard for AI management systems. For a company that faces customer audits, that fact matters more than one benchmark point.

The Meta case: an open licence is not a commitment

For years Meta was the main promoter of open models with its Llama family. Two facts remove it from that position.

First, Llama never had a free licence. The Llama 4 licence requires companies with more than 700 million monthly active users to ask Meta for permission, and its use policy does not grant the rights to the multimodal models to individuals domiciled in the European Union or to companies with their principal place of business there.

Second, in April 2026 Meta presented Muse Spark, a proprietary, cloud-only model with no downloadable weights. The company said the current Llama models will remain available; it did not commit to future versions. In August it published Muse Glimmer, a small model with 30 billion parameters under Apache 2.0, distilled from the closed model.

The lesson is about planning. Anyone who built on Llama has a model that still works, but no upgrade path. An open licence protects the version you already downloaded, not the next one.

The hyperscalers’ “sovereign” cloud

The other US offer is to keep the cloud and change its corporate governance. On 15 January 2026 Amazon launched its AWS European Sovereign Cloud in Brandenburg, with an investment plan of more than EUR 7.8 billion, German subsidiaries and operation exclusively by European Union residents.

European providers do not accept this as sovereignty. In March 2026, 24 executives from member companies of their association, CISPE, wrote to the European Commission; its secretary general asked the Commission to avoid “legitimizing sovereignty-washing”. They argue that the definition must be based on effective ownership and on protection from extraterritorial laws such as the CLOUD Act, not on having offices in Europe. The three US clouds together hold close to 70% of the regional market.

For Mexico the discussion is the same, with fewer local alternatives. The cloud regions that Microsoft, Google and Amazon opened in Querétaro between 2024 and 2025 belong to providers under that jurisdiction.

Policy: export the full stack

In March 2026 the Department of Commerce opened a new phase of the American AI Exports Program. It promotes “full-stack AI technology packages”: AI-optimized hardware, data center storage, models, cybersecurity and applications, with priority in export licence review for approved consortia.

This is Washington’s response to other countries’ demand for sovereignty: sell them the entire stack. For a buyer country, that solves capacity and leaves dependence intact.

What it means for a plant in the Bajío

• The useful range is already free. Gemma 4, gpt-oss and Granite 4.1 carry the Apache 2.0 licence. The first two fit on one 80 GB GPU according to their makers; Granite, at up to 30 billion parameters, is in the same range. You do not need the frontier model to classify documents, assist maintenance or query manuals in Spanish.
• A downloaded model does not phone home. It runs on the plant’s own equipment. The CLOUD Act targets providers that hold third-party data. A model that you operate on your own server does not put your data in anyone else’s hands.
• Plan the replacement. The Meta case shows that the next version may not arrive. Design the application so that you can change the model.
• If the customer audits, look at the certification. In automotive and aerospace supply chains, model traceability will be an audit question.

In this series: Guide: what sovereign AI is · 1. Aleph Alpha · 2. Mistral · 3. United States · 4. China · 5. Spanish-language models and Mexico

Sources: Google, Gemma 4, Apr 2026 · Google, Gemma terms of use · OpenAI, gpt-oss, Aug 2025 · OpenAI, gpt-oss and safeguard models · IBM Research, Granite 4.1, Apr 2026 · IBM, Granite 4.0, Oct 2025 · Nvidia, Nemotron · Microsoft, Phi-4-reasoning-vision · Meta, Llama 4 licence · Meta, Llama 4 use policy · The New Stack, Muse Spark, Apr 2026 · InfoQ, Muse Glimmer, Aug 2026 · Amazon, AWS European Sovereign Cloud, Jan 2026 · The Register, CISPE, Mar 2026 · US Department of Commerce, American AI Exports Program, Mar 2026 · Microsoft, Mexico Central region, May 2024 · Google Cloud, Querétaro region, Dec 2024 · DCD, AWS region in Mexico, Jan 2025

Bajío Ventures — a Mexican energy-infrastructure investment group based in Guanajuato: capital, a data platform on the region’s industrial base, and a German industrial channel, with its own advisory presence in Berlin.