Why Foundation Model Companies Are Pushing Agent Products Even While Losing Money

The signal truly worth heeding is not that AI companies are growing too slowly, but that they are growing fast while cash-flow pressure keeps mounting.

Author: Koutian Wu; GitHub: ktwu01

A lot of people ask: why are OpenAI, Anthropic, and xAI still investing heavily, and even pushing heavier Agent products? The answer is not simply “they’re greedy,” but that the industry is still in the “seize the structural position first, worry about profits later” phase.

Data First: Growing Fast, but Investing Just as Ferociously

Let’s start with facts supported by public sources (as of 2026-03):

On the OpenAI side, Reuters reports its CFO disclosed that 2025 annualized revenue surpassed $20 billion (vs. $6 billion in 2024), while compute capacity rose from 0.6GW in 2024 to 1.9GW in 2025. The meaning of this combination is direct: revenue is rising, but dependence on compute supply is expanding in parallel.

On the Anthropic side, its official announcement shows a $30 billion Series G completed in 2026-02 at a $380 billion post-money valuation; the same announcement gives run-rate revenue at $14 billion. It is not stagnating; it is, under high growth, continuing to raise to expand infrastructure and product capability.

On the xAI side, public coverage shows it completed a $20 billion funding round in early 2026 and continues to push capital into data centers and compute build-out. This shows it currently resembles a capex-driven race more than an enterprise in a “profit harvest” phase.

Specific loss numbers like the “$14B loss” currently come mainly from media reports of internal projection documents (e.g. The Information, relayed by Reuters) and should be treated as “sourced estimates,” not audited company financial statements.

Why “Google / Meta Can Make Money” Does Not Mean “Large Models Inherently Make Money”

Google’s and Meta’s public financials offer a very clear control group:

Alphabet 2025 revenue of $402.8 billion, operating profit of $129 billion, and capex of $91.4 billion; within that, Google Services operating profit was $139.4 billion and Google Cloud operating profit was $13.9 billion. They can absorb heavy investment because their core business cash machines are strong enough.

Meta 2025 revenue of $200.966 billion, operating profit of $83.276 billion, and capex of $72.22 billion; in the same year Reality Labs posted an operating loss of $19.193 billion. In other words, Meta uses its FoA profit pool to continuously subsidize a heavy-investment frontier direction.

So “top tech companies have profits” cannot be simply extrapolated to “standalone foundation model companies have entered steady-state profitability.”

Why Push Agents: This Is a Business-Structure Move, Not Product Sentiment

The key value of Agents is not “being cooler”; it’s converting one-off consumption into continuous consumption.

A single Q&A is usually low-frequency and short-funnel; but once you hook in search, browser, tool calls, and long-flow automation, the frequency, duration, and context depth of model invocation all rise. That directly changes per-user token consumption and retention structure.

In other words, an Agent functions more like a “usage-density amplifier”: it pushes the model from an answer-answering tool toward workflow infrastructure. For vendors, that is the decisive battleground for future pricing power and renewal quality.

This Path Also Has Boundaries: Heavy Investment Does Not Automatically Equal High Profit

To state it completely: pushing Agents is not a profit tonic. It merely gives the platform a path more likely to run long-term commercialization.

So the more accurate formulation is not “because we’re losing money, therefore we push Agents”; rather, “in the high-burn race, whoever seizes the high-frequency workflow gateway first has a better chance of converting today’s losses into future structural returns.”

This is also one of the underlying reasons a product like OpenClaw catches fire easily: it pushes AI from the “chat box” into a “sustainably executable operation layer.”

If you want a full product-level analysis, see my main article: OpenClaw: Turning the Command-Line World Into Your Second Brain.

Sources

• Reuters (OpenAI annualized revenue and compute capacity disclosure): OpenAI CFO says annualized revenue crosses $20 billion in 2025
• Anthropic official announcement (Series G, valuation, run-rate): Anthropic raises $30 billion in Series G funding
• Reuters (xAI funding): Musk’s xAI raises $20 billion in upsized Series E funding round
• Reuters relay (source of OpenAI 2026 loss projection): OpenAI tops $25 billion in annualized revenue, The Information reports
• The Information (original media source of OpenAI loss projection): OpenAI Projections Imply Losses Tripling to $14 Billion in 2026
• Alphabet 2025 Form 10-K (revenue, operating profit, segment profit, CapEx): SEC filing
• Meta Q4/FY2025 results (revenue, operating profit, Reality Labs loss, CapEx guidance): SEC Exhibit 99.1