Inside the Infron and AWS partnership

AWS is our infrastructure partner. On June 17, 2026, AWS ran its Hong Kong Summit at the Hong Kong Convention and Exhibition Centre under two words: Go build. One track that day was the AWS Executive Podcast, a full day of live recording where AWS hosts sat down with twelve executives for twenty minutes each. Andrew Zheng, our CTO, was one of the twelve. The AWS Summit Hong Kong page still lists him, under Certain AI, the company behind Infron.

AWS then recorded a second session, a 25 minute studio interview at their Hong Kong office. Two recordings, both produced by AWS, both on the record.

We are not a content company, so instead of describing the relationship, here are four clips from those two recordings, and then the part that actually affects you: what running on AWS changes inside the gateway.

Quotes below are lightly edited for readability. Nothing was changed in meaning. The clips are unedited.

AWS did not open with a softball. Their host went straight at the thing a cloud provider actually wants to know:

"You are an AI company building on AWS. Why not run your own applications on premises, and why would you select AWS as your cloud service provider?"

Yuna, who is founding GTM, answered on their stage:

"We are more focused on the AI solutions. And actually, we partner with AWS. AWS gives us comprehensive solutions, global scale, the reliability and availability of AI infrastructure. The second answer is, we are an AI startup. Running infrastructure is not an easy thing. It is very heavy. We want to focus more on our innovative solutions, not on running the infrastructure."

There is a version of Infron that spends two years standing up its own footprint in five regions and then tells customers about it. That company has better slides. It does not have a better router.

A gateway earns its keep on three things: which model your request should go to, what that costs, and whether your security team will sign off on where it ran. Owning the racks improves none of them.

A router is only as global as the map underneath it

The follow-up was about going global:

"What has AWS helped you most with in other global regions?"

Andrew:

"AWS is a global cloud company. That is the main reason we chose AWS as one of our most important partners and providers, because AI is a global business. In Europe, in the US market, there are so many startups and so many founders. They raise money and put AI into their new system stacks. So we need a global architecture like AWS, so we can run our business on it and deliver real value to our customers."

In practice this shows up earlier in a deal than people expect. The first question in an enterprise security review is not what a model costs. It is where the request gets processed, and who can read it. Answering that per request, instead of once per company, is a footprint problem before it is a software problem.

What the partnership changes inside the gateway

Andrew, on what the routing actually does:

"Ten percent of your traffic we route to a frontier model, like Claude in AWS. Other traffic goes to open source models, or to Nova. That is how we hold cost as a fixed line."

The reason he keeps coming back to a fixed line is a shape problem he hears from founders every week:

"The token is linear. When my business grows, my cost grows also. My revenue is like a line, but my cost grows. So I am losing money when my customers grow."

Here is the concrete version. Claude reaches you through Amazon Bedrock, Nova is Bedrock only, and both sit behind the same API as every other model on Infron. Choosing between them is two fields, not an integration. Pin the lane per request with the provider field, and the response names the provider it actually used, so the pin is something you verify rather than something you trust. On August 2, 2026 the Bedrock lane and the Anthropic direct lane both list claude-sonnet-5 at $2.00 per million input and $10.00 per million output, and both report zero data retention on the live catalog. The point is not the price, it is that the two lanes cost you the same amount of engineering.

The part we did not write

"For the next six to twelve months, I'm sure that AWS, we are going to embark on this journey together. We really look forward to working more together."

That is the AWS host closing her own interview. We did not write it, and we could not have.

What this is not

We are not claiming an AWS Partner Network tier in this post. What is on the record is what you just watched: an invitation to a stage AWS controls, two recordings AWS produced, and a Bedrock lane you can call this afternoon.

We are also not exclusive to AWS, and you should not want us to be. On August 2, 2026 the Infron catalog carried 464 models across 86 provider endpoints. A router with one road is not a router. The AWS lane is valuable because you can choose it, and because you can prove you chose it.

FAQ

Does using Infron mean my traffic runs on AWS? No. By default a request goes to whichever healthy provider serves that model best on price. If you want AWS specifically, pin it per request with the provider field. The same field keeps a workload off a provider you have ruled out.

Is Claude cheaper through Bedrock on Infron than direct? Not on list price. On August 2, 2026 both lanes list at $2.00 and $10.00 per million tokens for claude-sonnet-5. What differs is which lane your compliance team signs off on, and which one is healthy when the other is not.

Can I put Infron usage on my AWS bill? Not yet. We are working through the AWS Marketplace listing process. If AWS committed spend is how your budget actually works, say so, it changes our order of operations.

Where do I watch the rest? These four clips are the AWS parts of two longer recordings. The AWS Summit Hong Kong page has the event, the agenda and the list of twelve Executive Podcast speakers.

Try it on your own workload

Point your base URL at Infron, pin amazon-bedrock on one route, and compare it against whatever you run today. That takes an afternoon and it answers the question better than this post does. Start here.

Less orchestration. More innovation.