Alphabet Inc. has entered a precarious phase of its AI ascent—one that calls to mind a steel magnate who, while building new mills, finds his best foremen departing for a rival’s forge, even as he is forced to buy back his own metal at premium prices. The company now faces a dual threat: an accelerating exodus of top AI talent to Anthropic, and a severe internal compute deficit that has pushed it into costly, unconventional arrangements with SpaceX 6,11,14,15,16,21,44,47,51,54,67. These are not isolated troubles; they reflect deeper strains in Alphabet’s vertical integration strategy—strains that, if unchecked, could erode its command of the AI value chain.
How We Got Here: The Industrial Logic of AI
Throughout history, the masters of new industries have been those who controlled the critical raw materials—be it iron ore in the age of steel, or crude oil in the age of petroleum. In the AI era, the raw materials are compute cycles and data, but the true scarcest resource is human capital: the researchers and engineers who can refine algorithms into breakthrough capabilities. Alphabet, with its DeepMind arm, long held a commanding position, analogous to a trust that owned the mines, the railroads, and the processing plants. Yet that position is now under siege from within and without. The company finds itself provisioning a fast-rising rival, Anthropic, with the very advanced compute assets its own teams desperately need, while a stream of its most decorated researchers walk out the door to join that same competitor 11,12,13,14,15,16,17,21,29,33,47,49,53,55,56,62,66,67.
The Talent Exodus: The Flight of the Master Craftsmen
The talent drain from Google DeepMind to Anthropic has reached a critical velocity. In a single six-day period, at least four senior members departed 68. The most emblematic loss is that of John Jumper, a Nobel laureate, Vice President, and engineering fellow who left after nine years—an event reported by multiple sources and amplified by a 5% drop in Alphabet’s stock 11,12,13,14,15,16,17,19,21,24,28,30,33,39,47,48,49,53,55,56,58,59,60,62,63,64,66,67. He was quickly followed by Gemini researchers Jonas Adler and Alexander Pritzel 37,46,62. This exodus is not merely a handful of individuals; it represents an erosion of the institutional knowledge and inventive capacity that underpins Alphabet’s AI leadership 32.
The causes are becoming visible. Internal frictions over compute allocation have surfaced: shortly before one departure, compute from a project was reassigned to a London-based team 50,68. Alphabet has openly acknowledged that compute capacity constraints are throttling research output 29,50. In any industry, when a master artisan finds their tools stripped and their work stalled, the lure of a well-resourced competitor becomes decisive. Anthropic, by offering both abundant compute and a clear mission, has become a magnet for such talent.
The Compute Conundrum: Capacity Shortfalls and the Unlikely Alliance with SpaceX
For all of Alphabet’s massive infrastructure, its demand—from internal services, DeepMind, and Google Cloud—has outrun supply 3,29. This has manifested in throttling even large partners like Meta 25, imposing per-prompt limits on Gemini 2, and persistent difficulty in securing specific GPU types 23. In an extraordinary measure, Alphabet has turned to SpaceX, a vertically integrated competitor through xAI and Grok, to lease AI compute capacity 5.
The terms of these leases are startling. Alphabet pays $920 million per month for access to roughly half the capacity Anthropic receives at the Colossus 1 data center 6,44,51,54,65. Anthropic pays $1.25 billion per month 51, but on a per-GPU basis, Alphabet’s rate is 2.17 times higher 5. The combined deals involve some 435,000 NVIDIA GPUs 5 and generate revenue for SpaceX equivalent to 128% of its projected 2025 revenue 20. In short, Alphabet is paying a steep premium to a rival for the essential fuel of its AI ambitions.
This situation is the modern equivalent of a steel baron who, having failed to expand his own mines fast enough, must buy iron ore from a competitor—and at a price that enriches that competitor’s capacity to refine and sell finished steel. The capital outflow not only pressures Alphabet’s margins but directly strengthens xAI’s vertical integration, which now spans physical infrastructure, model-tier leasing, and developer tooling through the acquisition of Cursor 51.
The Deep Ambivalence of the Google–Anthropic Partnership
The relationship with Anthropic was envisioned as a classic cloud-provider-to-startup tie-up, but it has mutated into a complex web of cooperation and contention. Alphabet furnishes Anthropic with cloud services, custom TPU chips, and large-scale compute resources 27, and even financial backing for data center leases 27,38. Yet Anthropic competes directly with Alphabet in coding tools, frontier models, and enterprise solutions 4,36. Worse, Anthropic secured TPU capacity commitments from Alphabet before the latter fully grasped its own internal demand for Gemini development 29; Alphabet now sells Anthropic critical, often non-fungible compute that its own researchers need 29.
Anthropic, for its part, is wisely diversifying its compute supply. It has committed up to five gigawatts of AWS Trainium capacity 29,61, leased from SpaceX 4,52, explored space-based data centers with SpaceX 4, and partnered with Fluidstack for bespoke infrastructure 52. These moves reduce Anthropic’s long-term dependency on Alphabet, even as Alphabet benefits from near-term revenue. The enterprise traction of Anthropic—with alliances like KPMG 7, Micron 18,35,43,57, and the California state government 22,40,45—demonstrates that it is successfully commercializing its models, often on Alphabet’s own cloud rails 41,42.
Strategic Implications: Risks and Prescriptions
These intertwined dynamics introduce tangible risks to Alphabet’s investment narrative. Compute costs are ballooning, talent retention is faltering, and the perceived moat in AI research is blurring. If top researchers continue to depart, Alphabet’s ability to maintain parity in next-generation model capabilities could falter, ultimately impacting cloud growth and the advertising-driven AI features that rely on those models.
The path forward demands a return to the virtues of vertical discipline. Alphabet must accelerate its own infrastructure buildout 1,10,26 to eliminate the dependency on SpaceX and regain control over its cost base. Investments like the Antigravity platform, AI Hypercomputer, and custom Axion CPUs and TPUs 8,9,31,34 are steps in the right direction but must be pressed with greater urgency. More critically, the talent hemorrhage must be stanched—not merely with compensation, but by ensuring that researchers have assured access to abundant compute and a clear path to impact. Finally, Alphabet must candidly assess the boundaries of its cloud supplier role. Selling scarce TPU cycles to a direct rival may bring revenue, but at what cost to long-term strategic position? As in any industry, the master of the mill must decide whether he is a commodity provider or an empire builder. The present course is one of uneasy compromise; clarity of purpose will determine whether Alphabet remains the preeminent force in AI or cedes ground to hungrier, more integrated rivals.