AI Hardware: Digest for July 25, 2026

The AI race is hitting a wall of physics: computing power requires so much energy and capital that even giants like Google are recording historic financial dips. Meanwhile, AMD is preparing its answer to NVIDIA, and the US Army is running into token limits.

Google: First Negative Cash Flow in the Name of AI

Alphabet has reported its first-ever negative quarterly cash flow. The cause: unprecedented capital expenditures on AI infrastructure. While the company continues to post record revenues, investments in data centers and accelerators are now outpacing operating income.

In my view, this is a tipping point for the entire industry. The "infrastructure now, monetization later" business model only works with infinite access to cheap capital. If Google cannot convert AI spending into proportional revenue growth within the next 2–3 quarters, we will see a CapEx contraction across the sector.

Simultaneously, Google announced Gemini 3.6 Flash and confirmed the training of Gemini 4. New models require even more compute—meaning costs will only rise, not stabilize.

The Energy Ceiling: New York Depends on Canada

The 545-kilometer Champlain Hudson Power Express transmission line from Quebec to Queens has become critical infrastructure for New York. During July's heatwave, the state imported 52 GWh of electricity from Canada—9% of its total daily consumption.

This isn't just a local New York problem. AI data centers demand gargantuan amounts of energy, and the existing US grid infrastructure cannot keep up. While the CHPE project has faced delays, its current performance proves one thing: without a massive overhaul of power grids, the growth of AI compute will physically stop. Countries without excess power capacity simply won't be able to compete in the AI race.

AMD Advancing AI 2026: Betting on an NVIDIA Alternative

AMD is holding its Advancing AI 2026 conference to present new hardware and software solutions. The specifics are expected during the keynote, but the context is clear: the market is desperate for an alternative to NVIDIA accelerators, and AMD is the only real competitor in the data center segment.

From my perspective, AMD's primary hurdle isn't the hardware, but the software stack. NVIDIA's CUDA remains the de facto standard, and until AMD ensures seamless compatibility, even superior chips will struggle. However, given current AI infrastructure costs, buyers are willing to try any alternative to reduce vendor lock-in.

US Army Exhausts AI Token Limits

The US Army has hit an unexpected snag: troops were notified that their annual supply of AI tokens is rapidly running out. "Unlimited" plans turned out to be limited.

This is a symptom of a broader issue. AI compute does not scale infinitely, even with a budget. Infrastructure limits—compute capacity, network bandwidth, API access—are becoming the bottleneck. Anthropic, with its Opus 5 model, is betting on token efficiency rather than a leap in capabilities. This is the right signal: the era of "pay for any amount of compute" is ending; the era of optimization has begun.

Bottom Line

AI infrastructure is reaching a breaking point. Google is showing that even the biggest players cannot infinitely scale CapEx without a return on investment. Power grids are buckling under the load of data centers. Token limits are revealing the boundaries of available compute. From here, the industry either pivots toward efficient architectures and alternative hardware vendors, or the AI bubble collapses under the weight of its own costs. The next two quarters will reveal who can actually monetize AI and who was simply burning capital.