AWS、アプリケーション事業へ再参入し「墓石」を増やす

On Tuesday in San Francisco at an event called "What's Next with AWS," CEO Matt Garman took the stage to announce that AWS is (for what, depending on how you count, is the seventh, eighth, or ninth time) moving up the stack and entering the applications business.

The batting lineup:

• Amazon Quick, formerly known as QuickSight, formerly known as Q Developer, formerly known as Q Business, formerly known as Q. This feels pretty damned close to being the fourth rebrand in roughly eighteen months. The press release announced that it was "generally available today." Garman, on the same stage, said the desktop app is "in preview." Amazon Quick's own VP, in his keynote, simply said they were "launching new technology that will change how work gets done," not bothering to specify whether that meant GA or preview. Pick a lane. Please?

• Amazon Connect Decisions, a supply-chain planning product. You may know its competitors: Workday, o9, Blue Yonder, Kinaxis, SAP.

• Amazon Connect Talent, an AI-driven hiring product. You may know its competitors: Workday, SAP SuccessFactors, Greenhouse, Eightfold, iCIMS.

• Amazon Connect Health, a clinical AI product. You may know its dominant incumbent: Epic, with 40-percent-plus EHR market share, plus deep ambient partnerships with Abridge, Ambience, and Microsoft's Nuance.

• Amazon Connect Customer, the rebrand of the original Amazon Connect contact-center product. Existing customers who log in this morning will find their console renamed without consultation. On stage, AWS framed this as "the birth of a new family.' Whether the existing Connect customer base reads it that way is a separate question.

• And the headline-grabber: Bedrock Managed Agents powered by OpenAI, plus Codex on AWS, plus a multi-year deal involving AWS training capacity for OpenAI. Sam Altman appeared via pre-recorded video, having (in his own words) had his schedule "taken away from me today"—an unusually candid way of describing the Musk lawsuit court date that ate his afternoon. The video itself ran roughly five minutes, was clearly a one-take, and featured the OpenAI CEO stumbling over the teleprompter on at least two occasions while announcing the partnership AWS had just spent the previous half-hour treating as a marquee product reveal. The vibe was "I have a court date in twenty minutes; record it once, send the file." Whose partnership is this exactly?

Got all that? Good. Now we can talk.

A brief tour of the application graveyard

To understand why this matters, it helps to remember AWS's track record on applications: that is, software intended to be used by humans, rather than infrastructure intended to be used by engineers building software for humans.

A non-exhaustive tour:

• Amazon WorkMail: announced January 2015, end-of-life March 31, 2027. A decade-long answer to a question nobody asked.

• Amazon WorkDocs: launched July 2014 as Zocalo, renamed January 2015, full shutdown April 25, 2025. The phrase "WorkDocs Drive" still produces a small flinch in anyone who encountered it.

• Amazon Honeycode: launched June 2020, beta ended February 29, 2024. Less than four years on the shelf.

• Amazon WorkLink: launched January 2019, full shutdown November 30, 2021. Two years and eleven months, cradle to casket.

• Amazon Chime: launched February 2017, end of support February 20, 2026. Quietly phased out, then loudly stopped.

• The original Amazon Q: announced November 2023, it's shockingly difficult to enumerate the version history without a whiteboard. Now reincarnated as Quick Suite (October 2025), which absorbed Q Business and rebranded QuickSight in the process.

• Amazon QuickSight: announced October 2015, GA November 2016, rebadged into Quick Suite October 9, 2025. Never visibly threatened Tableau or Power BI; now technically also doesn't exist under its own name.

The one application AWS shipped, kept shipping, and that customers actually bought at scale is Amazon Connect, the contact-center product. Aubrey rightly noted on stage that Connect handles 20 million customer interactions a day and processed 12 million AI minutes (which are presumably like regular minutes only far more expensive) last year. Connect is good. Connect's customers are happy. Connect works.

It's worth asking why.

Why Connect succeeded and the rest didn't

A contact center is, fundamentally, infrastructure dressed up as an application. It's a routing-and-recording system with phones and dashboards on top. Routing is a thing AWS is excellent at. Recording is a thing AWS is excellent at. Once you frame "phone calls are packet streams that need to be reliably handled and stored," it becomes the kind of problem AWS can throw infrastructure muscle at and win.

WorkMail was different. Calendaring, mail rules, Outlook integration, address-book sync—those require you to know things about how humans actually work, not just how to scale an SMTP server. Microsoft has fifteen thousand person-years invested in Outlook. AWS launched WorkMail with the AWS console as its admin UI. The outcome was highly predictable and very depressing.

The pattern repeats: Honeycode required application-builder taste of the kind Glide and Retool spent years developing, not the taste of Amazon's website. WorkDocs required you to out-Google Google Docs. Chime required you to out-Zoom Zoom. Every time, AWS brought infrastructure-grade muscle to problems that demanded application-grade taste, and every time, AWS underperformed. The exception was Connect, because contact centers were infrastructure all along.

The same bet in nicer wrapping

Connect Decisions wants to compete with twenty- and thirty-year-old domain leaders in supply-chain planning. Its press release headline customer metric is "$85K of excess inventory savings in the first quarter," at a company called Wells Vehicle Electronics. For context: SAP, o9, Blue Yonder, and Kinaxis routinely cite seven- and eight-figure savings per deployment. $85K is, in supply-chain terms, well within the range of "the cost of running the planning meeting exceeded the savings." On stage, Wells and the only other named Decisions customer, TVS Motors, were described as "partners" who are "bringing the product into meetings and into discussions"—candid framing that reads as design-partner language. Y'know: pre-revenue.

Connect Talent's product page—the only one of the four new Connect SKUs that's actually live at this writing—says "Now available in Preview." It says nothing, in the section describing AI-led voice interviews, about New York City's Local Law 144, the EEOC's 2023 AI hiring guidance, or the EU AI Act's high-risk-system rules. For an AI hiring product launching in 2026, that's a conspicuous absence.

In fairness, the launch positioning makes clear that Connect Talent is aimed at high-volume hiring—Amazon's own seasonal warehouse staffing scale, plus retail, hospitality, customer support, and logistics. This isn't aimed at AI-screening engineering candidates, and it would be unfair to characterize it that way. It is, however, aimed precisely at the volumes where bias amplifies most efficiently, and precisely at the scenario Local Law 144 was written to address. "We're only doing this at scale" is not a get-out-of-compliance argument; it is the reason compliance posture matters most.

The page does promise that "every recommendation comes with clear, job-related reasoning and consistent evaluation standards, so when you make a hiring decision, you can demonstrate best practices and help your organization comply with hiring requirements." Translation: the customer is responsible for compliance. The vendor isn't volunteering anything specific. That's a landmine waiting for someone to stumble onto, and their terms of service make it very clear that it's the customer's responsibility, not AWS's. Connect Health appeared on stage with exactly one launch partner (NetSmart, a behavioral-health and post-acute IT vendor) and a pre-recorded video of physicians complaining about documentation burden. If you're a health-system CIO who watched the keynote expecting an Epic integration story, you saw what wasn't said.

Who, exactly, is the customer?

The strongest workflows demonstrated on stage all starred an Amazon employee.

The Quick desktop demo walked through prepping a customer meeting—pulling roadmap files "from my machine," account-team emails, internal Slack threads. That's an AWS field rep prepping for a customer call. Connect Decisions's pitch leaned heavily on "Amazon's 30 years of supply-chain experience" and the team that manages 400 million SKUs. Connect Talent's heritage story is Amazon's own seasonal hiring scale. Connect Health's flagship deployment is One Medical, an Amazon subsidiary. The Garman/White fireside's strongest agentic-development anecdote was Amazon's Prime Video team rewriting an internal payments system in two quarters with a handful of people.

This is consistent with the explicit framing on stage: AWS built these tools to run Amazon's own business, and is now offering them to customers. That is an honest origin story. It is also the same starting point as WorkMail and Honeycode and Chime: a tool built for Amazon's own use, projected outward, in the hope that the rest of the market wants the same thing, only to discover that it absolutely doesn't.

Whether AWS can capture external markets on its own with this generation of products remains to be seen. The demos, for now, suggest that the strongest customer is also the vendor.

And then there's "humorphism"

On stage, Connect's new design philosophy was introduced as humorphism: a coined word framed as the "dynamics of human interaction" answer to skeuomorphism's "physics of objects." This is, on its merits, an analogy that does not survive contact with daylight. Skeuomorphism described a real industry-wide visual translation pattern with decades of adoption. Humorphism describes one company's UX intent and asks the analogy to ride along. In keeping with AWS naming style, it starts with the root word "humor", yet is completely unfunny.

Whoever named this will, with luck, be promoted to a role with no naming responsibilities.

And meanwhile, Anthropic

I would be remiss not to mention this. AWS invested several/many/a horrifying number of billions of dollars in Anthropic. Today's keynote anchored on a brand-new opinionated agent SKU for OpenAI, with Codex on AWS, with AWS training capacity for OpenAI, and with the aforementioned five-minute Altman teleprompter outtake.

Anthropic's "moment," in case you missed it—and you probably did—arrived one week earlier in the form of a quietly published AWS Machine Learning blog post announcing that Claude Cowork and Claude Code Desktop now run on Amazon Bedrock. No keynote, no Dario video, just a blog. I've launched various shitposts with more fanfare.

AWS made a multi-billion-dollar Anthropic investment, and treated the corresponding Bedrock launch as a Tuesday blog post. AWS announced an OpenAI partnership eight weeks in the making, gave it the keynote anchor, and got back a five-minute video from a CEO who plainly had somewhere better to be. This could be deliberate news management. It could also be that Anthropic got the rolling-moment treatment because OpenAI is the partner whose models AWS customers are loudly demanding. Either read is interesting, but neither makes it particularly easy for the AWS marketing team to hold the "we are the place you can run both leading frontier models" line with a straight face for the remainder of the year.

The kicker

AWS is excellent at infrastructure. AWS is also, increasingly, a hundred-billion-dollar-plus business that believes its scale should substitute for domain depth in adjacent markets. The historical evidence says it cannot. Connect was the exception, because contact centers were infrastructure all along. Today's bet—the four-product Connect family, the third Q rebrand in eighteen months, the OpenAI anchor—extends the Connect brand across three new buying centers and assumes the bet generalizes.

I will sign up for Quick Desktop today, because it might genuinely be useful, and the early adopters (some of whom I even trust) claim it is. I'll give it a fair shake and tell you what I find. But the betting line on this entire launch should be priced against twenty years of AWS-application history. The graveyard is full. The new tenants have impressive nameplates.

Place your wagers accordingly. ®