By Akhull Khullar, investment principal and co-head private markets research at Arbra Partners
The AI sector has now officially moved beyond its initial hype phase. Sophisticated investors have seen a once-in-a-generation structural shift where revenues have demonstrably reached the foundational LLM layer in clear, quantifiable terms. OpenAI attained approximately $25bn in annualised revenue by early 2026. Anthropic accelerated even further, reaching a $30bn annualised run rate by April 2026.
These figures derive from real enterprise and developer spending on API access, subscriptions, and custom models, rather than internal credits or projections. Nvidia’s own leadership in terms of chip stocks and data centres, and the hyperscalers’ AI cloud businesses reinforce the picture that the foundational layer has delivered measurable ROI.
The central question has also become more nuanced. With infrastructure bets validating at the model layer, where will the next wave of returns accrue? Will a select group of application-layer companies capture the majority? Might enterprises realise sustained margin expansion through productivity gains and new revenue streams? Or could value emerge through some other channel?
This panoply of queries represents something of a reset, akin to the cloud transition of the 2010s, but are now even more compressed in time and elevated in scale. Capital that entered the infrastructure must now migrate towards a display of durable, compounding opportunities.
The revenue reality: Proof, not promise
By mid-2025 OpenAI had already exceeded its earlier $12.7bn full-year projection. Since then, execution has continued to outpace forecasts. Enterprise adoption now drives the majority of spend, with Anthropic securing a rising share of corporate LLM budgets.
This revenue reflects genuine customer commitments that have advanced well beyond pilot stages, and the economics begin to justify the preceding years of substantial capital expenditure.
See also: Electricity: A bottleneck or catalyst for AI?
Identifying the subtle signals, however, remain essential to the sector. OpenAI continues to forecast notable losses extending into the later part of the decade and hyperscalers anticipate annual AI infrastructure outlays measured in hundreds of billions.
The South African-American, entrepreneur-investor (and former COO of PayPal) David Sacks, has recently underscored a related dynamic specific to Anthropic. The company has sustained hyper-growth at a scale and velocity rarely seen even among the largest technology firms, conferring a valuation premium. Sacks has outlined a credible path for Anthropic toward a trillion-dollar valuation, driven by the rarity of such sustained momentum and the potential for duopoly-like positioning in frontier AI.
Application layer or enterprise margins?
The history of cloud application provides some precedent, with early infrastructure investment establishing the foundational pipes. The greatest compounding returns ultimately accrued to applications that leveraged them at scale and to the enterprises that reconfigured their operations around them.
Those that integrate AI as a core transformation driver are distinguishing themselves through superior revenue growth, innovation rates, and cost discipline.
See also: Mercer finds AI now used by majority of asset managers in investment process
Still, the data remains uneven. Many organisations observe productivity improvements yet encounter difficulty linking them directly to profit-and-loss statements. Leading adopters, meanwhile, are pulling ahead. The next wave of returns is likely to divide between two primary sources.
First, vertical and horizontal application-layer companies that embed AI so deeply into workflows that they become indispensable. These are AI-native platforms that did not exist prior to 2023. Second, enterprise customers that translate AI capabilities into tangible margin expansion through reduced time on routine tasks, elevated output, and novel revenue from AI-enhanced products and services.
Proof has advanced beyond anecdote, especially in those industries and areas where AI has become integrated into their whole ecosystems.
The blind spot: Portfolio repricing
Much of the capital deployed today continues to pursue infrastructure narratives or the most recent foundation-model entrant. Yet, discerning family offices and institutional investors have begun posing more demanding questions. They are looking for answers about which application layer companies will secure ownership of critical workflows? The enterprises that will demonstrate persistent margin expansion? And in which verticals (healthcare, finance, manufacturing) will AI compound into long lasting competitive advantages? Capital inflows into AI now respond to a concrete signal that the foundational layer functions, and well.
The ensuing revaluation will favour those positioned to capture downstream value, whether through specialised applications, agentic systems, or enterprises that have genuinely re-engineered core operations.
Today, the relevant inquiry for investors with extended time horizons is whether portfolios are structured around an AI environment where ROI accrues to the subsequent layer, and whether capital is being deployed with sufficient speed to secure it.
