I have been arguing for a while that the interesting money in AI is not the software line, it is the labour line — the trillions companies pay people to do work that is mostly a system with humans wedged into the gaps.6 That used to be a contrarian thing to say to a room of SaaS investors. It is not contrarian anymore. Over five weeks this summer the two largest cloud companies on earth stood up and said it back to me, with nine and ten zeroes attached.
So the honest question for anyone building an AI-native services company — which is what every venture I run at Gigaverse is — stopped being is this real and became did the giants just take it.
Four balance sheets, one bet
The framing that set the tone came from Sequoia in the spring. The partner Julien Bek's essay put it as plainly as anyone has: the next trillion-dollar company will not sell software tools, it will sell the work itself — a software company disguised as a services business, chasing the roughly six dollars companies pay people for every one dollar they spend on software.1 The category even has a market now: agentic software grew into a rough $9–10 billion global market in 2026, with forecasts reaching into the tens of billions by the end of the decade.1
Then the capital showed up. On 30 June, AWS announced a billion-dollar Forward Deployed Engineering organisation — teams of five or six engineers embedded directly inside a customer for roughly forty-five-day engagements, building and shipping production agentic systems alongside the client's own staff, and funded entirely off Amazon's balance sheet.2 It was not even first; TechCrunch noted AWS was following OpenAI and Anthropic, both of which had already stood up their own forward-deployed ventures earlier in the year.3 Forty-eight hours after AWS, Microsoft went bigger: a $2.5 billion "Frontier Company" with six thousand industry and engineering experts embedded with customers to, in its own words, "co-design, co-innovate, deploy and continuously improve AI systems at scale based on measurable business outcomes."4
Strip the logos away and all four are making the same wager: that the value in AI is not access to a model, it is someone standing inside your business making the model actually do the job. That is the whole services-as-software thesis, now underwritten by the four richest balance sheets in the field.
What the giants are actually buying
The first thing this does is end an argument. Gartner had already projected that 40% of enterprise applications would ship with task-specific AI agents by the end of 2026, up from under 5% a year earlier.5 When AWS and Microsoft commit thousands of humans to the last mile, they are conceding the thing that took me longest to convince people of: the model is not the product, the deployment is. You do not spend $2.5 billion on six thousand people to hand out API keys.
The second thing it does is tell you exactly which customers they want. Look at who they named. AWS went live with the Allen Institute, Cox Automotive, the NBA, the NFL, Ricoh and Southwest Airlines.2 Microsoft opened with the London Stock Exchange Group, Land O'Lakes, Unilever and Novo Nordisk, with Accenture, Capgemini, EY, KPMG and PwC as delivery partners.4 These are the largest enterprises in the world, served through the largest consultancies in the world. That is not an accident of a launch press release. It is the only shape of customer that pays back a forty-five-day engagement staffed by a hyperscaler's own engineers.
A billion-dollar forward-deployed org is not built to fill a dental chair or clean a mid-market catalogue. It is built for the customers whose logos justify the org.
Which means the land grab is real, and it is happening entirely at the top of the market. The tail is a different business, and they are not in it.
Where the model leaves margin on the table
Read the AWS design goals closely and you find the tell. Their model is explicitly built so the customer is self-sufficient when the deployment ends — agentic-first, timelines compressed from months to days, engineers out the door in roughly forty-five.2 That is a genuinely good product for a Fortune 500 with its own platform team waiting to take the handoff. It is also a decision to walk away from the part of services-as-software that actually compounds.
Because the enterprise reality I keep running into is that go-live is not the finish line. The old deploy-and-maintain motion becomes deploy-and-continuously-re-engineer as a customer adds product lines, edge cases and workflows, and the reliability bar climbs from the 80% that closes a pilot to the 99%-plus that survives production — a last stretch that can cost a hundred times the effort of everything before it.6 A forty-five-day engagement is aimed at the cheap 80%. The expensive, durable part — the operating layer that keeps the thing at 99% while the business changes underneath it — is precisely what a "self-sufficient at handoff" model hands back to the customer.
The tail the land grab can't reach
Here is the market the four balance sheets are structurally unable to serve. A single dental practice trying to capture every missed call and fill every chair. A mid-market Shopify merchant whose catalogue needs to be legible to the agents that now do the recommending. A young company that needs formation, banking and compliance handled end-to-end rather than explained. None of these is a Fortune 500. None of them justifies a hyperscaler embedding six engineers for six weeks. And there are tens of thousands of each.
That tail does not want a self-sufficient handoff. It has no platform team to hand off to. It wants the outcome owned on an ongoing basis by someone who has done this specific unglamorous thing before — which is the exact opposite of the giants' model, and the exact thing a focused studio is built to do. The expensive machinery of doing it well — the ingestion kits, the telemetry, the deployment scaffolding, the senior team that has shipped it in a neighbouring vertical — transfers across verticals even though the domain knowledge does not. That is the whole reason I run Gigabit, GigaCommerce, Top Dentistry and FormBridge off one senior team instead of four separate companies.6
I can make this less abstract without naming a client. The last mile, in practice, is almost never the interesting-sounding part. It is discovering that a dental office's "schedule" is three calendars that quietly disagree, that the front desk has an unwritten rule about which patients get squeezed in, and that the real job is not a smarter model but capturing the call, reconciling the calendars, and booking into the one that actually governs the chair. It is the compliance step that only fires when a filing crosses a threshold nobody wrote down. None of that survives a demo, and none of it is visible until you are standing inside the operation.
That is precisely the work a forty-five-day engagement is built to hand back — and it is the work I want to own, because once it is encoded it keeps paying. The messy specifics are the moat, not the obstacle.
What this means if you are building one
The three sentences I would give a founder. Do not try to out-enterprise the hyperscalers — they will win the London Stock Exchange and you will not, and that is fine, because the logo you are chasing does not fit in their model anyway. Pick a vertical below their radar where the service is a system with people in the gaps, and where the buyer has no platform team to catch a handoff. And build for the right of the handoff line: own the operating layer that keeps the outcome at production quality while the customer's business changes, because that is the part that compounds and the part they have chosen to give away.
The land grab is not the threat it looks like on the day the billion-dollar headlines land. It is the biggest marketing budget the category has ever had, spent convincing every enterprise on earth that paying for outcomes instead of tools is normal now. The hard part was never getting people to believe services become software. AWS and Microsoft just retired that problem for all of us. The hard part is, and remains, the last mile — and in the tail of the market, the giants are not standing in it.
One honest caveat, because the symmetry is too clean otherwise. "The giants can't serve the tail" is a claim about today's economics, not a law of nature. The same tooling that makes a hyperscaler's forty-five-day engagement viable — agentic-first scaffolding, reusable ingestion, self-serve handoff — is exactly what could push their cost-to-serve down the market over time, and their partner armies at Accenture and PwC already reach further than their own engineers do.4 The studio bet is that focus and a compounding operating layer stay ahead of that descent in verticals too small and too weird to templatise. I believe that bet. I would not pretend it is risk-free.