Newsletter · · Ashutosh Agarwal

Hyperscalers and AI Labs Are Quietly Becoming Your Consultant - IT Services vs AI - Week of September 19, 2026

IT Services vs AI for the week of September 19, 2026: a channel analyst put dollar figures on hyperscalers and AI labs dropping 500,000-dollar-a-year forward-deployed engineers into clients (Microsoft $2.5 billion, AWS $1 billion, Google $750 million), Accenture bought roughly 580 SAP and CRM specialists in 48 hours for its new Accenture Edge mid-market unit, Ramp data showed the heaviest enterprise AI spenders cutting per-employee spend 10% in August, and Accenture's fiscal Q4 print is set for October 1.

IT Services vs AI

Week of September 19, 2026: Hyperscalers and AI Labs Are Quietly Becoming Your Consultant

TL;DR

  • The disruption is now going top-down, and it has a price tag. Hyperscalers and AI labs are dropping their own $500,000-a-year engineers straight into client accounts (Microsoft has committed $2.5B, AWS $1B, Google $750M, just doubled, to "forward-deployed engineers") to become "the seventh-and-a-third partner in the room." That is the systems integrator getting cut out from the top, quantified for the first time. (Business of Tech, Sep 17)
  • Accenture is buying its way into the mid-market. It snapped up two SAP shops (Tokyo's Comware, ~180 people; Holland's McCoy, ~400) within two days in August and folded them into a new mid-market unit, "Accenture Edge." The pitch is "industrialized delivery": templates, accelerators, speed. (Transformation Ground Control, Sep 16)
  • Enterprise AI spending actually fell last month. The heaviest corporate AI spenders cut per-employee spend 10% in August, to $7,200 from an $8,000 July peak, a soft-demand signal that cuts against the "AI is a bottomless services tailwind" story. (The AI Daily Brief, Sep 18)

What's new

The five developments that most move the thesis this week, ranked by how much they matter to a book.

1. The disruptors are becoming the consultants, and now we have the dollar figures

This is the biggest development of the week, and it is the clearest evidence yet for a thread we have been chasing: the AI companies themselves are moving into services and cutting the integrator out from the top.

On Business of Tech (Sep 17), host Dave Sobel interviewed Jay McBain, one of the most-cited channel analysts in the industry, an operator/insider who has spent decades mapping how technology actually gets sold. His point: the hyperscalers and model labs are worried that their partners (the consultants and integrators) are too slow to turn all that AI infrastructure into real projects, so they are increasingly doing the consulting themselves.

In his words, they are "going to keep dropping in expensive resources. We're talking $500,000 a year people into the customer directly to become that [seventh]-and-a-third partner in the room to get to that outcome so that they can start telling investors to calm down." The commitments he cited to these "forward-deployed engineers": Microsoft $2.5 billion, AWS $1 billion, Google $750 million and just doubled it. Why does it matter? Because a "forward-deployed engineer" sitting inside a client, paid for by Microsoft or Google, is doing exactly the high-value, before-the-sale design and architecture work that an Accenture or an Infosys used to own.

McBain layered on several other numbers worth writing down:

  • The money to turn partners into consultants is enormous. Enablement spend this cycle: "Google spent $750 million. Anthropic spent $100 million. ChatGPT, OpenAI, $150 million. SAP, $100 million euros. Salesforce dropped $50 million." And "even Dell, Michael Dell himself, moved billions from the point of sale back into these consulting moments."
  • AI projects are starting to actually work. He referenced the widely-quoted failure rates, the MIT "95% of pilots stuck" headline from a year ago, and "the following week Accenture came and said, no, no, our failure rate's 90%." His update: "earlier this summer, we approached 70 percent. And by the end of this year, we're going to knock that number below 50." A falling failure rate is the bull case's oxygen: it means pilots convert into production, and production means recurring managed-services revenue.
  • But 80% of the growth is in enterprise, not mid-market. McBain's firm quantified the AI channel at "32.5% compounded growth… 80% of that's in enterprise." Down in small and mid-market, where a lot of integrators live, "it's not really happening at the same level."
  • The billing model is changing under everyone's feet. He expects MSP contracts to look like Microsoft's new "$99 a month plus" tiers, a base fee plus a variable, consumption-based, token pass-through charge. Hours-based billing is quietly being replaced by "you press the magic button, the bill could be $500 for that seat."

"They're going to do it themselves. They're going to keep dropping in expensive resources… $500,000 a year people into the customer directly." (Jay McBain, on hyperscalers and AI labs building their own consulting arms)

Why it matters: last week the "model labs moving into services" idea was a claim. This week it is a set of budget line items. For anyone short the integrators, this is the cleanest evidence that the squeeze is coming from above (labs and hyperscalers) as well as below (cheap agents). For the bulls, McBain's falling-failure-rate and post-2028 managed-services argument is the counter.

2. Accenture is buying the implementation labor pool and calling it "Accenture Edge"

On Transformation Ground Control (Sep 16), the best operator source we have on the economics of big-project delivery, host Eric Kimberling (founder of Third Stage Consulting) walked through Accenture's mid-market push in concrete detail:

"Accenture agreed to buy Tokyo-based a company called Comware in August, which is 180-ish SAP and CRM professionals based out of Tokyo, and this is just two days after they agreed to buy McCoy, which is a Dutch SAP partner with close to 400 SAP specialists… they're folding both of these companies into a business unit called Accenture Edge, which is a mid-market entity they launched in June."

That is roughly 580 SAP/CRM specialists acquired in a 48-hour window, all pointed at the mid-market. The stated pitch, per Kimberling, is "industrialized delivery… templates, accelerators and speed." But he flagged the catch that matters for the thesis:

"Those accelerators may compress the timeline, but they don't remove the client-owned work… A faster delivery engine will happily transport your old confusion, your old broken processes at speed." In other words, Accenture can automate its own side of an ERP project, but the part that actually makes projects slip, the client's own data governance and process decisions, doesn't go away. AI speeds up the vendor's work, not the whole implementation.

This dovetails with a second Accenture item from Strategy Simplified (Sep 16), an episode literally titled "Why Consulting Demand Is Outpacing Every Firm's Hiring Plan." The hosts (management-consulting-industry analysts) opened with Accenture's new partnership with Google Cloud "to sell prebuilt AI agents to [mid]-market companies in the… $300 million to $3 billion revenue range." Their read: "A lot of these tech providers are turning into distribution channels for consulting… it's high-margin revenue for them." They described the market as a dumbbell, big scaled firms on one end, hundreds of boutiques on the other, and "the firms in the middle are getting squeezed."

Why it matters: Accenture is playing offense, using AI partnerships (Google Cloud) and acquisitions (Comware, McCoy) to grab mid-market share and reposition as the distribution arm for hyperscaler AI. The bear reading is that "industrialized delivery" is a tacit admission that the old, people-heavy big-project model is under pressure, and that Accenture is buying up the very SAP-implementation labor pool that AI is supposed to shrink.

3. Enterprise AI spend fell 10% last month: the demand signal cuts against the tailwind story

On The AI Daily Brief (Sep 18), host Nathaniel Whittemore relayed the latest Ramp AI spending index (a card-spend dataset; skewed to tech-forward early adopters, so treat as directional). The finding, from Ramp lead economist Eric Karazian: "in August, the top 1% of businesses spent $7,200 per employee per month… down 10% from a July peak of $8,000."

Ramp's own explanation is not "demand is dying"; it is that the most sophisticated buyers are getting better at using cheaper models: "price cuts plus a growing share of spend is shifting to standard and light models which are already cheaper." (Whittemore added his own caveat that summer seasonality is probably underweighted.) Either way, the direction, the heaviest AI spenders cutting per-head spend, is a soft signal for anyone counting on AI budgets to fund an endless services boom.

The same episode carried a sharp build-vs-buy data point: Latham & Watkins, the second-largest US law firm, is reportedly buying NVIDIA servers to stand up its own in-house models as an alternative to OpenAI and Anthropic APIs. And Foundation Capital's Jaya Gupta argued every software company should "become a model factory for its own vertical." When large, sophisticated buyers build rather than hire out, that is demand walking away from the integrator.

Why it matters: this is the hardest demand data of the week (theme 5), and it leans bearish. It also confirms last week's softer signal (CIOs with "pressure to show impact but not a lot of new investment").

4. The zero-coding engineering org: an operator lives the headcount-decoupling thesis

On 20VC (Sep 12), Harry Stebbings interviewed Matteo Franceschetti, co-founder and CEO of Eight Sleep, an operator running a real company, not a pundit. His description of how AI has reshaped his org is the headcount-vs-revenue decoupling thesis made flesh:

  • "Our engineers stopped coding around a year ago. Since then, they didn't code. What they have is hundreds of AI [agents] that code for them." That, he said, is how a "fairly tiny team" operates across 35 countries including China and the Middle East.
  • 160 people total, with "revenue per employee… way higher than Apple," and output he pegged at "5x what is average in Silicon Valley." His finance team is 4 people where a comparable company would run 20; his email-marketing team went from two people to "a team of zero" running "close to 100 million."
  • Three-year plan: "250 people making a billion", where similar companies "are in the thousands."

Franceschetti also gave the single best framing of the coding-tools spend debate. Referencing Marc Benioff's comment that Salesforce spends "$300 million a year on Claude and they have a $6 billion budget for engineering" (i.e., AI coding tools are ~5% of the engineering budget), Stebbings asked whether that 5% goes to 25% (bullish for the labs) or down to 1% (commoditization). Franceschetti's answer: both vectors at once. "The usage will increase… the 5% will become 50%, but the cost will go down. And so net-net, I think it will go down," comparing it to how the cost of electricity became trivial over time.

Why it matters: the linear headcount-to-revenue model, the foundation of the traditional IT services and staffing pyramid, is being abandoned by operators in real time. If a growing company can hit a billion in revenue with 250 people and zero human coding, the bear case on labor-arbitrage-based services is not hypothetical.

5. The coding-agent arms race got repriced upward, but margins are a warning

Two episodes captured the money pouring into AI coding tools and the productivity claims behind it:

  • On The AI Files (Sep 12), the host reported that Cognition (maker of Devin) raised $2 billion at a $48 billion valuation, roughly double its $26 billion mark from four months earlier. He cited a company-reported revenue run rate growing "to $492 million to nearly $900 million since May." (Flag: that revenue figure is company-reported and unverified; treat with caution.) His takeaway: "AI coding remains far from winner-take-all," with Microsoft's GitHub Copilot and others still very much in the race.
  • On Marketing in the Age of AI (Sep 17), the host gave the productivity soundbite of the week, with a coding tool, "the thing that used to take a dev team a quarter now takes an afternoon by a prosumer", and the revenue-per-person example everyone cites: Midjourney at "about $200 million in ARR with roughly 11 employees." But he also delivered the crucial counterweight: AI companies "are watching their gross margins erode by six points or more just from what it costs to run the AI. Inference is eating about 23% of their revenue." The thin "wrapper" businesses have no moat.

Why it matters: the tools that displace billable engineering hours keep getting cheaper, better and better-funded: that is structurally bad for headcount-based services. But the eroding-margin point is a reminder that the disruptors' own economics are not bulletproof.

The debate

Bull: AI grows the services pie and lifts margins. Everyone is trying to build with AI, and most of them fail without help: the failure rate is falling from ~90% toward under 50% (McBain), which means a wave of pilots is about to convert into production and recurring, post-sale managed services by 2028–2031. GenAI is becoming a distribution channel into consultants, not around them: Accenture is Google Cloud's route to the mid-market (Strategy Simplified). Cycle times are collapsing (one AWS partner went from "six-month cycle times to four weeks" using GenAI, per Ultimate Partner, Sep 14), so a firm can simply run more projects. And someone still has to do the messy, client-owned work that AI can't touch: data governance, process decisions, change management (Kimberling). The title of this week's Strategy Simplified episode says it plainly: consulting demand is "outpacing every firm's hiring plan."

Bear: AI absorbs the billable work and breaks the model. The disruption is now coming from above as well as below: labs and hyperscalers are dropping $500K forward-deployed engineers straight into clients (McBain), doing the high-value design work integrators used to own. Enterprise AI spend is actually falling (Ramp, -10%). Sophisticated buyers are choosing to build rather than buy (Latham & Watkins going in-house). The headcount-to-revenue link is broken: operators are running global companies with zero human coding and 5x the output per head (Franceschetti). Pricing is shifting from hours to consumption and outcomes (McBain, Strategy Simplified), which deflates the revenue line even where the work stays. And the mid-market, the dumbbell's thin bar, is being crushed between scaled giants and cheap boutiques.

Stocks in play

Accenture (ACN): by far the most-discussed name this week

  • Bull: Accenture is on offense. The Accenture Edge mid-market unit, the Comware and McCoy acquisitions (~580 SAP/CRM specialists in 48 hours), the Google Cloud agent-distribution deal, and a separately-reported AI-safety partnership with Anthropic all show a firm using AI to expand reach, not retreat. Its scale and privileged access to executive teams make it the natural distribution partner for hyperscaler AI (Strategy Simplified, Transformation Ground Control).
  • Bear: "Industrialized delivery" is a euphemism for automating away billable hours; accelerators shrink the very engagements Accenture bills for. It is buying the SAP-implementation labor pool that AI is meant to compress, a bet that could age badly if agents disintermediate ERP configuration. And the forward-deployed-engineer trend (McBain) is Accenture's problem too: hyperscalers doing the consulting themselves.
  • Next catalyst / number to watch: fiscal Q4 2026 earnings, Wednesday Oct 1 (before market open), the print we have been waiting for. Watch new bookings, the GenAI bookings run-rate, headcount trajectory, and any commentary on AI-augmented pricing. Investor day follows on Oct 14. (Context, not from a podcast: sell-side positioning into the print skewed cautious this week, with several mixed-to-negative analyst actions; the bar has been lowered.)

IBM (IBM): no direct podcast coverage this week

No episode this window addressed IBM Consulting, watsonx, or Consulting Advantage directly, so the read-through applies in full: forward-deployed engineers, mid-market pressure, and consumption-pricing all bear on IBM Consulting the same way they bear on peers. Next catalyst: IBM's Q3 print in October: watch Consulting bookings and the watsonx/Consulting Advantage attach rate.

Infosys (INFY): no direct podcast coverage (9th straight week)

For the ninth consecutive week, no English-language podcast discussed Infosys (or any Indian IT major) by name, a structural blind spot in this medium. The read-through is bearish: the pool of offshore SAP/dev talent is being consolidated by the giants (Accenture's Comware/McCoy), forward-deployed engineers attack the high end, and the headcount-decoupling thesis (Franceschetti) directly threatens labor-arbitrage models. Next catalyst / number to watch: fresher-hiring numbers and the Q2 FY27 print: any sign of headcount growth decoupling from revenue.

Wipro (WIT): no direct podcast coverage

Same read-through as Infosys. The only company-specific item this window was a solution launch (Wipro STO360 with Saudi Aramco and SAP), sourced from news, not podcasts, and modest. Next catalyst: the next quarterly print and any large-deal/booking commentary.

Read-throughs

  • TCS, Cognizant, Capgemini, EPAM: the read-through is bearish for the same reasons as Infosys/Wipro, offshore-pool consolidation, forward-deployed engineers, and consumption pricing.
  • Enterprise software vendors (Salesforce/Agentforce, ServiceNow, Workday, SAP/Joule): the most important read-through this week. Per Jay McBain on Business of Tech, the mid-market and SMB now get their agentic AI through these SaaS platforms directly ("Salesforce, ServiceNow, Workday, HubSpot"), not through a systems integrator, and the billing is moving to base-plus-consumption. Transformation Ground Control noted SAP's acquisition of Reltio (AI-first master data management), with the takeaway that enterprise AI "is actually a data story more than… a true AI story": good for whoever owns the data layer, and a reminder that data cleanup is still services work. If AI implementation flows through the software vendor rather than the SI, that is a slow bleed of integration revenue.
  • Microsoft / GitHub Copilot: still named as a leading coding agent alongside Cognition's Devin (The AI Files); Microsoft is also among the biggest committers to forward-deployed engineers ($2.5B, per McBain).
  • In-house AI build-vs-buy: tipping toward build at the high end: Latham & Watkins buying NVIDIA servers, and the "every software company should be a model factory" argument (The AI Daily Brief). Every enterprise that builds is an enterprise that isn't hiring an integrator.

What changed vs last week

Last week (Sep 12) delivered a rich cluster on SI-economics disruption. Here is what moved:

  • The "model labs to services" thread got quantified. Last week it was a claim (OpenAI/Anthropic "building forward-deployed engineering arms with PE"). This week Jay McBain put real numbers on it: Microsoft $2.5B, AWS $1B, Google $750M (doubled) into forward-deployed engineers, and $500K/year people going directly into clients. Top-down disintermediation is now a budget line, not a rumor.
  • "Accenture Edge" went from a footnote to a strategy. Last week it surfaced only as a Guggenheim reference (a supposed sign of soft core demand). This week we got the actual M&A behind it, Comware and McCoy, and the "industrialized delivery" playbook, plus Kimberling's important nuance that accelerators don't remove client-owned work.
  • Soft demand got hard data. Last week's signal was qualitative (CIOs with "pressure to show impact but not a lot of new investments"). This week Ramp put a number on it: enterprise AI spend down 10% among the heaviest spenders.
  • Consumption / outcome-based pricing keeps spreading, confirmed again by both McBain (base-plus-token model) and Strategy Simplified (outcome-linked contracts as a differentiator). Consistent with last week's BCG "tectonic shift to outcomes."
  • The Accenture Q4 print we have been waiting on for weeks now has a firm date: Oct 1. It did not land this week.