Odyssey Raises $310M at $1.45B, OVHcloud Trains Frontier Models, and the Week Europe Remembered It's Allowed to Build AI

It was a quiet week for product launches — most of the noise came from funding rounds and infrastructure deals that suggest the next wave of AI builders are betting on hardware access over model differentiation.

Odyssey raises $310M and locks in AWS compute

AI lab Odyssey closed a $310M Series B at a $1.45B valuation this week, with a side deal that gives it access to AWS's specialized ML chips. The round size puts it in the same weight class as recent infrastructure-heavy bets (Atom Computing's $225M, Cortea's $100M), but the AWS partnership is the more interesting signal.

We don't know which chips — could be Trainium, could be the new MTIA-adjacent hardware AWS has been testing internally — but the partnership structure mirrors what Anthropic and a handful of others negotiated in 2024 and 2025. Odyssey gets guaranteed capacity; AWS gets a customer willing to build on non-NVIDIA silicon.

The company hasn't said much about what it's training, but $310M and a custom compute deal suggests they're not fine-tuning Llama-3. This is either a large multimodal model or a domain-specific system that needs serious scale.

OVHcloud is training frontier models and that matters more than it sounds

France's OVHcloud announced it's training frontier AI models from scratch, making it Europe's second player in the LLM space after Mistral. The phrasing — "frontier" — is deliberate. They're not talking about a fine-tuned retrieval model or a vertical assistant. They mean a GPT-4 / Claude-tier system trained on tens of thousands of GPUs.

This is the first time a European cloud provider has said it will train and host its own foundation model. Mistral is a model lab that happens to be French; OVHcloud is infrastructure that decided it needed its own weights.

The implications are less about model quality (we have no benchmarks yet) and more about sovereignty and cost structure. If OVHcloud can train a competitive LLM and offer inference at cost — no margin stacking between model lab, cloud provider, and API gateway — European enterprises get a viable third option that doesn't route through AWS, Google, or Azure.

It also puts pressure on AWS and Microsoft to keep European data in Europe, which has been a slow-moving fight since GDPR. A homegrown LLM trained on EU-resident compute with EU-domiciled hosting changes the procurement conversation for banks, healthcare systems, and governments.

SageMaker adds inline payloads for async inference

Amazon SageMaker AI Async Inference now supports inline request payloads, which means you can send inference data directly in the InvokeEndpointAsync API call instead of uploading to S3 first.

For payloads up to 1 MB, you skip the S3 round-trip. That cuts latency for smaller requests (image classification, short-form text generation, JSON-in / JSON-out workflows) and simplifies code — no pre-signed URLs, no S3 bucket permissions, no cleanup jobs.

It's a small feature but it removes a common pain point. Async inference is already the right tool for batch jobs, background tasks, and anything that doesn't need sub-second response times. Now it's also easier to use for teams that don't want to manage object storage lifecycles.

The 1 MB ceiling is reasonable — if your payload is larger than that, you're probably doing video, large documents, or multi-image batches, and S3 is already the right place to stage that data.

What we're watching

The Odyssey and OVHcloud announcements both point to the same shift: model training is becoming an infrastructure play. The companies raising nine figures this year aren't the ones with the best demo — they're the ones with access to compute, data pipelines, and distribution that doesn't depend on OpenAI or Anthropic staying friendly.

Europe specifically is making a bet that it can build competitive models and keep them on European silicon. If OVHcloud's LLM hits GPT-4 performance at half the API cost, AWS and Azure have a problem. If it doesn't, this becomes a footnote in the "Europe tried" narrative.

We'll know which story we're in by Q4.