Two days ago, Anthropic announced a $1.5 billion joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs — a forward-deployed services firm that embeds Anthropic engineers inside mid-market companies and competes head-to-head with the Big Four. The thesis was loud: consulting is broken, AI labs can do it better.
This morning, Anthropic told the same industry a very different story.
EPAM Systems — a $4.7 billion engineering services firm with 60,000+ employees — announced a strategic multi-year partnership with Anthropic that commits the company to certifying 10,000 Claude architects, including 250 forward-deployed "Black Belts", with 5,000 architects certified by end of Q3 2026 (up from 1,300 today) and 20,000 EPAM employees already enrolled in Anthropic Academy training. EPAM is not being replaced by Anthropic. EPAM is being re-platformed on Anthropic.
So which is it? Is Anthropic the consulting industry's executioner — or its preferred substrate? The honest answer is both, and the contradiction is not an accident. It is a deliberate two-track distribution strategy that, taken together, tells you more about how enterprise AI is going to be sold over the next three years than any single product launch this quarter.
This piece is about the EPAM deal itself, why it matters more than the press release suggests, how it fits with Anthropic's broader Partner Network play, and what CIOs evaluating their AI services strategy need to understand before their next vendor RFP.
Update — August 7, 2026: EPAM reported Q2 2026 results on August 6, and they settle two of the questions below — one in this article's favor, one against it. The certification pace beat the benchmark set here: more than 5,700 Anthropic-certified engineers, past the 5,000-by-Q3 milestone a quarter early, with the 10,000-architect target pulled forward to year-end. But the third CIO implication below — that the multi-vendor hedge is getting harder to maintain — was wrong, and EPAM is the counter-example. The same call disclosed 2,000+ Gemini-certified specialists and a first-year commitment to certify 5,000+ OpenAI consultants. The hedge widened. The commercial picture is the part that should change what you do: AI-native revenue passed $160 million, over 11% of the business, inside total revenue of $1.415 billion, up 4.5%, with full-year guidance cut to 3.2–4.2% and delivery headcount essentially flat. A fast-growing AI line inside a flat top line is substitution, not expansion — and that is leverage if you are renewing a services MSA this quarter.
What EPAM Actually Committed To
The certification numbers are the headline, and they are bigger than they look on first read.
10,000 Claude-certified architects. EPAM has roughly 60,000 engineers globally. Committing one in six of them to a single vendor's certification track is not a marketing partnership — it is a structural decision about which AI substrate EPAM will deliver against for the rest of the decade. That is the kind of capital-T commitment that makes other partnerships (with OpenAI, with Google) look like hedge positions.
250 Black Belts. This is the most interesting line in the press release, and it gets the least attention. Black Belts are specialized, forward-deployed engineers — the same operating-model term Palantir uses, and the same forward-deployed model the Anthropic-Blackstone JV is building. EPAM is creating an internal cadre that can ship into a Forbes Global 2000 customer and run a Claude implementation end-to-end. Twenty embedded engineers per region, on average. That is a serious investment in the implementation layer that most services firms are still trying to figure out.
1,300 architects certified at announcement, 5,000 by Q3, 10,000 long-term. The pace matters. Five thousand certifications in five months means EPAM has already built the training pipeline, the certification testing infrastructure, and the internal incentive structures to push engineers through. This is not a "we'll announce a goal and figure it out later" partnership. The training is shipping now.
That held. On the Q2 2026 earnings call on August 6, EPAM put the number at "more than 5,700 certified engineers already ahead of the 5,000 by Q3 milestone we set out" — a quarter early — described itself as among Anthropic's top five globally certified partners, and moved the 10,000 Claude-certified architect target to year-end 2026 rather than the vaguer long-term horizon of the original announcement.
20,000 EPAM employees in Anthropic Academy. That number is a tell. It means a third of EPAM's workforce has touched Anthropic-curated training material in the last several months. Even if only half of them go on to certify, EPAM will have more Claude-skilled engineers than every Big Four consultancy combined. As of the Q2 call, more than half the delivery organization has completed Anthropic Academy coursework.
The product surface is equally specific: EPAM will deliver against Claude models, Claude Code, Claude Agent SDK, and Claude Security. That last item — Claude Security, the application security capability Anthropic launched into public beta last week — is notable because it brings a defensive-AI offering into the partnership, not just productivity tooling.
The deal terms are not disclosed, which usually means the economics involve revenue-share, joint pipeline development, and bidirectional referral commitments. Translation: EPAM gets first-look on Anthropic's enterprise pipeline; Anthropic gets EPAM's services capacity for delivery. Both sides are betting that they need each other.
The Black Belt Program Is the Strategic Center
Most readers will skim past the "250 Black Belts" line. They should not.
Forward-deployed engineering is the operating model that has won every enterprise-AI implementation race I have studied. Palantir's FDE program is the canonical example: engineers embedded inside customer accounts, with end-to-end ownership of outcomes, billing through services rather than software-license revenue. The Anthropic-Blackstone JV is, structurally, a clone of that model. Cohere's enterprise services team uses the same pattern. JPMorgan's internal AI deployment relies on a similar model with internal-employee FDEs.
What EPAM's 250-Black-Belt commitment does is productize the forward-deployed engineer at services-firm scale. Where Anthropic-Blackstone might field 50 senior engineers across North America, EPAM can field 250 across every major enterprise market in Europe, North America, and India. That is a different scale of distribution.
For Forbes Global 2000 buyers, the implication is concrete: by Q4, you will be able to procure a Claude-native implementation team at the SOW economics of a services firm rather than the equity-investment economics of the Anthropic JV. The unit economics of services delivery, paired with the technical depth of vendor-certified architects, is a faster path to deployment than either pure-play consulting or pure-play vendor services.
The risk on EPAM's side: Black Belts are talent-bottlenecked. Twenty-five experienced engineers per region is a small number; if utilization spikes — and given current enterprise AI demand, it will — quality control will become the binding constraint. The 5,000-by-Q3 certification target is partly designed to feed the pipeline behind the Black Belts.
The Claude Partner Network Context
EPAM is not alone, and that matters for understanding Anthropic's distribution thesis.
When Anthropic launched the Claude Partner Network in March 2026, it committed $100 million for 2026 to partner enablement and named four anchor firms in the launch announcement: Accenture (training 30,000 professionals on Claude), Deloitte, Cognizant (350,000+ associates with Claude access), and Infosys. EPAM joining today brings the named partner roster to five major services firms representing more than half a million engineers with some level of Claude access or training.
The network includes:
- A formal Claude Certified Architect (Foundations) credential, with seller, architect, and developer tracks shipping later in 2026
- Anthropic Academy training materials and sales playbooks
- Dedicated Anthropic Applied AI engineers for live customer deals
- Technical architects for complex implementations
- Code Modernization starter kits for legacy migration projects
- A Services Partner Directory for enterprise-buyer discovery
- Co-marketing and joint pipeline development
The structural read is that Anthropic has decided not to be the channel — at least not exclusively. The Wall Street JV is one bet; the Partner Network is the other. Together, they cover both ends of the enterprise services barbell: the high-touch, equity-aligned forward-deployed services for mid-market, and the broad-distribution, services-firm-scaled delivery for the Global 2000.
OpenAI runs a similar two-track playbook. Its enterprise channel program added Accenture, Capgemini, CGI, Cognizant, Infosys, PwC, and Tata Consultancy Services as launch partners back in April for Codex distribution; The Development Company JV (also announced May 4) is OpenAI's forward-deployed services bet. The two-track strategy is now table stakes for frontier AI labs that want to win enterprise share at scale.
Why EPAM Picked This Bet
EPAM's competitive context tells you why the bet was inevitable.
EPAM's growth slowed in 2024–25 as the legacy "outsourced engineering" economics compressed. The company's Belarus / Ukraine engineering base is geopolitically exposed. Its core competitor in Eastern European delivery (Globant) has been moving aggressively into Google Cloud / Gemini partnerships, and TCS / Infosys / Cognizant have multi-vendor AI strategies that hedge across OpenAI, Anthropic, and hyperscaler models.
A 10,000-architect bet on a single vendor is a sharp, deliberate market positioning move: EPAM is choosing to be the premier Claude services firm rather than a competent multi-vendor services firm. The bet pays off if Anthropic continues to win mid-market and Global 2000 share through 2027. It loses badly if a frontier-model market reset narrows Claude's relative position, or if a Big Four firm wins a similar exclusive partnership and out-spends EPAM on certifications.
Three months on, that either/or was the wrong frame. EPAM ran the single-vendor depth play and the multi-vendor hedge simultaneously — see the correction under "What This Means for CIOs" below. The concentration risk the paragraph above describes is correspondingly smaller than it looked in May, and so is the differentiation.
EPAM Chief Strategy Officer Elaina Shekhter framed it bluntly: "EPAM is pioneering the next era of enterprise innovation as the premier architects of AI-native builds." Read past the marketing gloss and the strategy is clear: stop competing on engineering hours, start competing on AI-native implementation expertise, and price accordingly. The services-margin story for the next 24 months is being written in this kind of partnership announcement.
What This Means for CIOs
Three concrete implications.
One: Claude-certified is now a procurement filter. If you are RFPing AI implementation services, the certification credential is becoming a meaningful evaluation criterion — the way "AWS Certified Solutions Architect" became a filter for cloud migration RFPs in 2018–19. By end of 2026, every major services firm will publish certification-headcount numbers. The number that matters is not "we have 50 engineers familiar with Claude" — it is "we have N% of our delivery roster certified through Anthropic Academy." Add it to your RFP scoring rubric this quarter.
Two: forward-deployed engineering pricing is the new benchmark. EPAM's Black Belt program will, by year-end, be priced at a premium to traditional services-firm rates but at a discount to the Anthropic-Blackstone JV's equity-aligned model. That middle-tier pricing is the new market benchmark for productized AI implementation. If your incumbent services partner is bidding on AI projects without a forward-deployed offering, you are paying senior-consultant rates for project-management work — and you should expect a 20–30% pricing renegotiation.
Three: the multi-vendor hedge is getting harder to maintain. Services firms used to compete on tooling neutrality — "we'll deploy whatever foundation model you prefer." That posture is breaking down. EPAM's 10,000-architect bet on Claude means EPAM's billable-hour quality on a Claude project will be materially better than its quality on an OpenAI or Gemini project. By Q4, vendor specialization will start showing up in delivery quality, even when the firm publicly maintains multi-vendor capability. Buyers should ask: which AI vendor have you actually invested certification budget against, and what does that mean for delivery quality on my project?
Correction — August 7, 2026: this implication was wrong, and EPAM is the counter-example. On the Q2 call EPAM reported its 5,700+ Anthropic-certified engineers alongside more than 2,000 certified under Google's Gemini Certified Partner Specialist program — against a 5,000-by-Q3 target there too — plus a first-year commitment to certify more than 5,000 OpenAI consultants and train over 10,000 specialists, made when it joined OpenAI's Partner Network at the Advanced tier on July 28. EPAM did not pick a substrate and give up neutrality. It bought certification depth against three of them inside four months. The hedge widened; it did not narrow.
What survives is the closing question — but ask it with a sharper edge. Demand a number per vendor, and separate certified headcount today from commitments to certify: EPAM's Anthropic and Google figures are the former, its OpenAI figure is the latter, and only one of those can staff your project this quarter.
Where the Bet Is Fragile
Three risks worth tracking.
Concentration risk on Anthropic itself. A 10,000-architect bet only pays off if Anthropic remains a top-two enterprise frontier vendor through 2028. If the next Claude release misses materially, or if a competitor (Google Gemini, OpenAI GPT-6, an open-source frontier model) takes meaningful share of the enterprise stack, EPAM will have a large, expensively-trained workforce certified against an underperforming substrate. The recovery cost would be material.
Cannibalization by the Wall Street JV. Anthropic-Blackstone is, in some accounts, EPAM's competitor for the same mid-market AI implementation work. If the JV scales aggressively and Anthropic prioritizes its equity-aligned channel over its certified-partner channel, EPAM may find itself bidding into customer accounts where Anthropic's own JV has a structural advantage. The Partner Network terms presumably address this with channel-conflict protections, but the details are not public.
Certification commoditization. The Claude Certified Architect credential is currently scarce — 1,300 people globally hold it. By Q3 2026, with EPAM past 5,700 already and Accenture, Deloitte, Cognizant, and Infosys also pushing certifications through, the credential will commoditize — faster than this section assumed, since EPAM cleared its Q3 number in Q2. The premium pricing window for "Claude-certified" services is a 12–18-month window. EPAM needs to convert the head start into long-term customer relationships before the certification differential erodes.
What to Do Before Q3
Three actions for the buy side — plus a fourth that the Q2 numbers added.
First: redo your AI services RFP rubric this quarter to include certification-headcount metrics, forward-deployed engineering capability, and AI-vendor specialization disclosures. The vendors that cannot answer these questions cleanly are not the vendors you want for your 2026–27 AI implementation roadmap.
Second: assume the Anthropic two-track strategy is the template other frontier labs will follow. OpenAI is already there. Google Cloud's $750M agent partner fund (announced at Cloud Next '26) is the same play. By end of 2026, every major model vendor will run both a forward-deployed equity-aligned services arm and a certified-partner network. Plan your services procurement around that reality, not around the multi-vendor neutrality fiction of 2023–24.
Third: negotiate the certification-headcount disclosure into your existing master services agreements. Quarterly reporting on the percentage of your delivery team that holds active certifications across the relevant AI vendors is the kind of vendor-management discipline that prevents quality drift. It is also the lightest-touch lever you have to push your services partners toward the AI-native delivery model that will define the next decade of enterprise IT.
Fourth, added August 7: define "AI-native revenue" in the contract, and read what its shape tells you about the rate card. EPAM's $160 million figure appears in the earnings call, not in the audited Q2 press release — it is a management-defined, unaudited metric. If you are writing AI-native delivery commitments into an MSA on the strength of a vendor's disclosure, define the term yourself or you are buying a number your vendor gets to compute. Then look at the shape of it: AI-native revenue crossed 11% of the business and has grown double digits sequentially for six quarters, while total revenue rose only 4.5% to $1.415 billion, delivery headcount moved 0.3% from Q1 to roughly 56,650, utilization sat at 78.3%, and full-year guidance came down to 3.2–4.2% reported. That is substitution, not expansion — the AI-native work is displacing traditional billable engineering inside the same client base rather than adding to it. Substitution at flat headcount is deflation in the effective rate card, and it is your leverage, not theirs. The regional split says the same thing louder: the Americas grew 0.5% to $806 million against EMEA's 10.9% to $582 million, with management conceding North America "is not growing fast enough" and that it would operate below expectations in the second half. A vendor telling its investors that has less room to hold price than its proposal to you implies.
EPAM's bet is bold, specific, and structurally significant. Anthropic's two-track distribution strategy — replace consulting at the top end, re-platform consulting at the broad end — is the template the rest of the industry is going to copy. The question for every CIO is whether their incumbent services partner has placed a comparable bet, or is still hedging. The answer to that question is going to determine procurement outcomes through 2028.
Q2 answered it for EPAM: both. It placed the bet and kept hedging — and the market marked the stock down 13.7% to $94.83 on the day anyway, because the certifications are real and the growth they were meant to produce is not there yet. CEO Balazs Fejes put the thesis as "coding gets automated, engineering doesn't." He is probably right, and it is still not the number on your invoice. The coding is getting cheaper. Make sure your contract says so.
Want to calculate your own AI ROI? Try our AI ROI Calculator — takes 60 seconds and shows projected savings, payback period, and 3-year ROI.