In the second quarter of 2026, Alphabet reported negative quarterly free cash flow for the first time in its history. It burned $5.9 billion and then raised its full-year capital-expenditure forecast to between $195 billion and $205 billion.

The strange part is that this did not happen during a collapse in revenue. Google Cloud revenue rose 82 percent year over year, while Alphabet’s total revenue exceeded market expectations. Cash did not disappear because Google had stopped making money. It disappeared because Google was investing in AI infrastructure even faster than its enormously profitable businesses could generate cash.

The most obvious explanation is Gemini’s weakness. More precisely, the current controversy concerns the delayed Gemini 3.5 Pro, not a model called Gemini 3.7. Google released lighter models, including Gemini 3.6 Flash, but its flagship Pro model was delayed after reportedly falling short of internal objectives, particularly in coding and agentic work. The market is waiting for evidence that Google can still match the top systems produced by OpenAI and Anthropic.

Yet describing Alphabet’s cash burn as a desperate attempt to repair a failed model misses the larger story. Google is not merely paying an engineering bill. It is paying a war tax imposed by America’s AI religious civil war.

On one side stands the commercial accelerationist pole represented by OpenAI. OpenAI is not formally an e/acc organization, and it retains genuine safety policies. Its revealed behavior, however, is accelerationist: build larger models, acquire more compute, deploy rapidly, enter governments and enterprises, and make the advance of artificial intelligence economically and institutionally irreversible.

Within this worldview, the risks created by technology are not solved primarily by slowing development. They are solved by creating superior models, stronger technical safeguards and more capable institutions before competitors reach the frontier. AGI is therefore not only a danger. It is a revelation—a possible source of scientific abundance, productivity, military advantage, medical progress and economic transformation.

Data centers become cathedrals, accelerators become relics and training runs become rituals intended to bring the promised future closer. Hundreds of billions of dollars in infrastructure spending can then be interpreted not as financial excess but as the sacrifice required to reach the next stage of civilization first.

On the opposing side stands the safety doctrine most clearly institutionalized by Anthropic and influenced by the Effective Altruism and longtermist ecosystem.

Effective Altruism began in Oxford as an attempt to use evidence and cost-effectiveness to do as much good as possible with limited resources. Its early intellectual image was the comparison of measurable interventions: how many deaths could be prevented by spending a dollar on malaria prevention rather than on a less effective charity?

Longtermism expanded the time horizon of this moral calculation. It argues that efforts to improve the world should be evaluated largely by their effects over thousands, millions or even billions of years. Once vast numbers of potential future people are included in the calculation, even a small reduction in the probability of human extinction may appear to outweigh enormous improvements in present-day welfare.

At this point, Effective Altruism acquires an eschatological dimension. The effectiveness of malaria nets can be tested against real-world mortality. The expected value of preventing a hypothetical superintelligence from extinguishing trillions of potential future lives cannot be measured in the same way. A very small probability is multiplied by an almost unlimited future value, allowing speculative catastrophes to dominate present resource allocation.

Anthropic has embedded this worldview into corporate governance. It is a public benefit corporation whose stated purpose is the responsible development of advanced AI for humanity’s long-term benefit. Its Long-Term Benefit Trust was designed to exercise influence over the composition of the board without sharing the ordinary financial interests of investors. Its Responsible Scaling Policy links increasing model capabilities to escalating security and safety measures intended to address catastrophic risks.

Calling this a religion is intentionally provocative, but the metaphor is not arbitrary. Anthropic speaks in the language of humanity’s long-term future, catastrophic risk, transformative technology, institutional mission and moral responsibility toward generations that do not yet exist. Its mission is not treated as an ordinary corporate objective; the company describes it as the final arbiter of organizational decisions.

The conflict between OpenAI and Anthropic should not, however, be reduced to rationalists fighting madmen, or heroes fighting villains. The two camps sell different forms of apocalypse.

The accelerationist apocalypse says: If we slow down, a rival—especially China—will reach the future first and determine its rules.

The longtermist apocalypse says: If anyone reaches the future without adequate control, humanity may permanently lose control of its own destiny.

One doctrine treats delay as catastrophe. The other treats acceleration as catastrophe. One locates salvation in speed; the other locates salvation in control.

For a time, the U.S. government had reasons to use both. In 2025, OpenAI and Anthropic each offered broad government access to their enterprise AI systems for a nominal price of one dollar. The government needed AI both as a productivity tool for civilian agencies and as a strategic capability for intelligence and national security.

By early 2026, however, this unstable balance had broken down. Anthropic insisted on maintaining restrictions involving mass domestic surveillance and fully autonomous weapons. The Trump administration argued that lawful military use should ultimately be defined by the state rather than by a private technology company. President Trump then ordered federal agencies to phase out Anthropic’s technology, while the Pentagon reached a separate classified-deployment agreement with OpenAI.

OpenAI also retained red lines concerning domestic mass surveillance, autonomous weapons and certain automated high-stakes decisions. The difference was that it successfully translated those principles into contractual and technical arrangements the government accepted.

This confrontation reveals the true constitutional question behind the AI safety debate. The issue is not whether safety matters; every major laboratory claims that it does. The question is who possesses legitimate authority to define acceptable risk.

Is it the company that built the model? The elected government? The military? Investors? Independent trustees? Or a technical elite that claims to represent the interests of future humanity?

The Trump administration’s public position is explicitly geopolitical. It describes AI as a race against adversaries and argues that American companies must be free to innovate without burdensome regulation. From that perspective, extreme AI safety demands can resemble unilateral technological disarmament. If China will not pause, a unilateral American slowdown may be interpreted not as moral responsibility but as a strategy for losing national power.

Google is trapped between these doctrines.

Unlike OpenAI, Google is not a startup with little existing business to protect. Unlike Anthropic, it cannot define itself primarily as a safety-first mission organization. Google owns a vast established order built around advertising, search, cloud computing, Android, YouTube and productivity software. It must participate if AI transforms that order, yet it cannot allow a competitor to become the interface through which users access information, software and digital labor.

Google’s spending is therefore not a conventional investment made solely because management has calculated an attractive return on capital. It is also an insurance premium against strategic irrelevance. It buys compute, infrastructure, talent and time. Even if the precise economic returns remain uncertain, the potential destruction of Google’s existing franchise if it falls behind may be too large to tolerate.

In ordinary corporate finance, a company should invest when expected returns exceed its cost of capital. In an AI arms race, however, one participant’s spending forces the others to spend. OpenAI’s expansion compels Google to expand. Google’s expansion pressures Anthropic, Microsoft, Amazon, Meta and Chinese laboratories to mobilize still more resources.

No participant has fully demonstrated the ultimate return on this investment, but none can safely be the first to stop. It is simultaneously a technological race, a military-style arms race and a prisoner’s dilemma intensified by quasi-religious conviction.

Alphabet’s negative free cash flow therefore signifies more than the disappointing progress of a single Gemini model. It marks the temporary suspension of ordinary financial discipline under the pressure of a civilizational narrative.

OpenAI promises salvation through abundance. Anthropic promises salvation from extinction. Competition with China supplies both religions with the hard fuel of national security.

Google may not wholly believe either doctrine. It must nevertheless participate in the world those doctrines have created.

Even unbelievers must pay the tithe.

Alphabet’s $5.9 billion of negative free cash flow was that tithe.