AI Demand and Deflation

AI Demand and Deflation

Artificial intelligence is at once EPAM's largest source of new demand and the most direct threat to what it actually sells — billable engineering hours. The company has scaled a defined "AI-native" revenue line to a target above $600M for 2026 and now calls itself a "pure-play AI-native" firm, while its own 10-K warns that AI could force price concessions and says the fear has "negatively impacted the price of our stock" [1]. On the evidence to date the tailwind is winning at the top line; the pricing threat is real but not yet visible in the operating numbers.

This is the assumption underneath every estimate in the Financials and Estimates chapter: the estimates are most sensitive to which side of AI proves larger, and to whether that leaves mid-single-digit organic growth a floor or a ceiling.

The risk, in the company's own words

The threat is not analyst conjecture; EPAM has been escalating it in its filings for two years. Generative AI did not appear as a demand-and-pricing risk in the FY2021 or FY2022 10-Ks at all. It entered with the FY2023 filing, which warned that clients "may be unwilling to pay rates for human delivery personnel if they perceive that the same services can be performed by" AI, and that the technology "could disrupt … our ability to charge for their services" [2]. That first appearance coincides with the second leg of the share-price decline traced in Fallen Star.

By the FY2025 10-K the language had hardened into a dedicated section — "Risks Related to Artificial Intelligence" — with a sub-heading that states the fear plainly: "Increased Adoption of AI-Based Software Tools May Reduce Demand for Our Services." The filing concedes that clients "may seek other service providers or expect price concessions" if comparable work can be done "less expensively using AI," and, unusually for risk-factor boilerplate, admits the market has already reacted: competition from "AI-based task-specific tools … has negatively impacted the price of our stock" [3].

The mechanism is specific to EPAM's model. It bills largely on a time-and-materials basis: revenue is roughly hours multiplied by rate. If AI makes an engineer materially more productive, the same deliverable takes fewer hours, and under an unchanged pricing model that saving flows to the client, not to EPAM. The company gave a live example this year: at consumer-finance client Nelnet, its AI tooling drove a "31% productivity increase" and "nearly 2x" faster back-end development [4]. Delivered under pure T&M, a 31% productivity gain is a 31% headwind to the revenue on that scope.

The tailwind EPAM is selling against it

Management's answer is that AI enlarges the pie faster than it deflates the per-hour price. At its March 2026 Investor Day, EPAM framed a total AI-spending market of $4.7 trillion by 2029, growing at a roughly 30% compound rate, and defined its own addressable "AI Services" slice at about $1.3 trillion [5]. The thesis, repeated across recent calls, is that AI raises enterprise complexity and technical debt faster than it removes work — that "the build opportunity is a long-term win" as the backlog of required modernization becomes more evident, not less [6].

Unlike most tailwind stories, this one now carries a revenue line. EPAM reports a "pure AI-native" revenue figure — defined narrowly as AI-native IP and platforms plus AI-led transformation programs, explicitly excluding ordinary work merely accelerated by AI tools. That figure exceeded $105M in the fourth quarter of 2025, and management guides it above $600M for 2026 [7]. Against consensus 2026 revenue near $5.73B, a $600M AI-native line is on the order of 10% of the company.

The scale of the ambition is easier to read as the two disclosed points and their share of the business:

Source: Q4 FY2025 earnings call, prepared remarks [8]. A Q4 quarterly figure and a full-year target are not directly comparable; shown as the two points management disclosed.

The demand signals behind the number are broad rather than one or two marquee deals. On the Q1 2026 call, EPAM said more than 80% of its top-100 clients are engaged in AI initiatives, that it launched over 100 new AI-native projects in the quarter, and that it sees "an escalating large-deal pipeline focused on AI-enabled vendor consolidations" — multiyear deals larger than its historical norm, whose "full potential … is not yet reflected in our outlook" [9]. The pivot has a history: the "AI/Run" delivery framework and "AI factory" construct date to late 2024 under then-CEO Arkadiy Dobkin [10], and have since been made the center of the strategy under CEO Balazs Fejes.

The deflation test in the numbers

If AI were already hollowing out a labor-based business, the tell would be revenue rising while headcount falls, or revenue per employee climbing sharply as machines replace juniors. EPAM's disclosures show neither yet. Total employees fell to 53,150 in 2023 during the post-invasion contraction, then grew to 61,200 in 2024 and 62,850 in 2025 [11]. Revenue per employee has been range-bound, not deflating.

Source: employee counts from FY2025 10-K, Human Capital [12]; revenue per employee derived from reported revenue and year-end headcount, FY2022–FY2025. Reported revenue includes acquisitions, which lift the 2025 figure.

The read is genuinely two-sided and worth stating carefully. That headcount is still rising cuts against the crude "AI replaces the engineers" bear case — as of end-2025 there is no evidence in the workforce data that AI is shrinking the labor model. But the flat-to-modest revenue per employee also means AI has not yet delivered the productivity leverage the bull case needs; the 2025 uptick is partly the higher-value NEORIS and First Derivative acquisitions rather than organic mix. And management is explicit that the commercial model is mid-transition. On the Q1 2026 call, Fejes described "a generally new and consequential commercial construct," said the company will "evolve our approach to AI investment pricing … for some quarters to come," and pointed to a deliberate shift toward "new fixed-price and other service deals" and outcome-based models — the mechanism by which a services firm keeps a share of the productivity gain rather than handing all of it back [13]. Whether that repricing holds is not something the numbers can yet answer.

The whole industry is in the same bind

EPAM is not facing a company-specific disruption; every services peer is managing the same tension, which is a useful check on how EPAM is positioned. The clearest statement of the pricing dilemma came from Globant's CEO Martín Migoya, asked how "AI Pods" — a token-based, "more work with tokens and less with humans" model then at roughly $21M of annualized recurring revenue, under 1% of revenue — avoid cannibalizing seat-based revenue: "I am not in a position to prevent cannibalization. So I want that transformation to happen" [14]. The instinct across the group is the same: lead the cannibalization rather than be a victim of it.

Scale, though, separates the field, and here EPAM is a credible specialist rather than a leader. Accenture — the closest large-cap comparator named repeatedly in EPAM's own competitive framing — reported "Advanced AI" bookings of $2.2B in a single quarter (Q1 FY2026), nearly double a year earlier [15]. Cognizant said it had crossed $10B of cumulative AI-led large-deal starts, including $1.2B in one quarter [16]. EPAM's disclosed AI figure — a $600M revenue target, not a bookings tally — is proportionally meaningful to a company its size but an order of magnitude smaller in absolute terms.

Sources: Accenture Q1 FY2026 call [17]; Cognizant Q4 FY2025 call [18]; EPAM Q4 FY2025 call [19]; Globant Q4 FY2025 call [20]. Each firm defines and measures its "AI" line differently — bookings versus revenue versus recurring revenue — so the figures rank scale, not like-for-like performance.

The honest conclusion from the peer set is that EPAM's "pure-play AI-native" branding is a positioning choice, not a demonstrated share lead. Its edge, if it has one, is depth: independent validation this year included a Gartner Leader placement in generative-AI consulting and a top-five ranking for its AIRON developer agent on the SWE-bench software-engineering benchmark [21]. Whether premium engineering depth outcompetes the raw scale of Accenture and the India majors is a moat question this chapter does not settle.

The calibrated read

The weight of the evidence is that, so far, AI is adding more to EPAM's demand than it is subtracting from its pricing: a defined AI-native revenue line scaling toward roughly a tenth of the company, broad client engagement, and organic growth that re-accelerated to 5.6% by the fourth quarter of 2025 [22] rather than collapsing as a pure-deflation thesis would predict.

The strongest fact against that read is that the pricing risk is structurally real and largely unpriced by EPAM itself: the company is still delivering 31% client productivity gains under a commercial model it admits is only mid-transition [23], and its Q1 2026 outlook already flagged softer North America demand and "lower visibility in the second half" as clients defer larger discretionary programs [24]. The AI-native figures are also company-defined and unaudited, and the deflation could arrive with a lag — showing up in bill rates and margins before it shows up in headcount.

What would change the read in either direction is concrete and near-term: whether the AI-native line actually clears $600M in 2026 with stable or better margin (tailwind confirmed), or whether revenue per employee and organic growth stall even as the AI-native line grows (deflation winning, the top line masking a mix shift). The margin path is where the two forces will first collide — the subject the Financials and Estimates chapter leaves open.