NVIDIA’s latest earnings hit like a sledgehammer to Wall Street’s nervous system—$81.6 billion in quarterly revenue, up 85% year-over-year, with guidance for another 95% surge next quarter. This isn’t growth; it’s a full-throttle sprint. And trailing just behind aren’t mere followers but two trillion-dollar behemoths with burning eyes and deeper agendas: Google and Meta.
Don’t be fooled by the word “partnership.” The relationship among these three resembles Cold War nuclear powers—shaking hands in public while stockpiling arsenals in secret. NVIDIA supplies the ammunition (H100s, B100s, GB200s), while Google and Meta build AI armories in their backyards. They buy NVIDIA chips by the truckload to fill data centers, yet simultaneously refine their own custom silicon: Google’s TPU v5e, Meta’s MTIA v2. Do they truly trust Jensen Huang? No. They’re simply not ready to flip the table—yet.
I believe this so-called “AI triumvirate alliance” is a fragile truce of convenience. NVIDIA’s dominance rests on the CUDA moat, but that moat is eroding. Google’s Pathways architecture and Meta’s PyTorch-plus-MTIA stack are deliberate attempts to bypass CUDA’s lock-in. They don’t need to replace NVIDIA overnight—they just need leverage when Blackwell Ultra supply tightens or U.S. export controls suddenly clamp down.
Even more telling is how capital markets react. Meta is reportedly prepping an IPO for its “AI subsidiary”—a brilliant move. Packaging AI infrastructure as a standalone entity eases valuation pressure and proves to investors it’s more than just a social network. Meanwhile, Google’s I/O showcase of Gemini Live and Project Astra may seem consumer-facing, but it’s really a message to enterprises: our AI isn’t just a chatbot—it’s an OS-level layer woven into workflows. Both are transforming AI from a cost center into a profit engine, albeit through different paths.
But here’s the fatal question: has infrastructure investment already wildly outpaced real-world AI applications? Data center power consumption soars, NVIDIA chip lead times stretch to months, and TSMC’s CoWoS capacity sells out instantly. Yet beyond ad optimization and recommendation engines, how many AI use cases actually generate steady cash flow? Healthcare, manufacturing, finance—AI adoption there remains sluggish. It echoes the dot-com bubble: fiber optics laid across oceans, but no one knew how to monetize them.
NVIDIA knows the risk. That’s why Jensen keeps hammering the “AI factory” narrative—positioning GPUs as the industrial looms of a new era. Clever, yes, but perilous. If markets realize these “factories” churn out models and data with no clear path to revenue, confidence could evaporate overnight.
Will Google and Meta sit idle until then? I doubt it. They’re hedging with dual strategies: publicly embracing NVIDIA to stabilize the ecosystem, while privately accelerating in-house chip and software integration. Meta’s open-source Llama series isn’t just about community—it’s building an alternative to closed, proprietary models. Google, through Android and Chrome, pushes AI to the edge, reducing absolute reliance on cloud compute.
The endgame may not be about winners and losers, but a redistribution of computational power. NVIDIA might remain the biggest beneficiary, but its monopoly premium is thinning. Google and Meta are playing a longer game—they don’t just want better chips; they want to define the operating systems and protocol layers of the AI age.
So while everyone cheers the $81.6 billion headline, I’m watching the unreported details: MTIA yield rates in Meta’s fabs, TPU cluster utilization at Google, the tired eyes of NVIDIA engineers debugging next-gen NVLink at 3 a.m. Those are the real signals shaping AI’s next three years.
One last thought: as the triumvirate feasts on centralized compute, has anyone considered that AI’s true breakthrough might not come from the cloud—but from the edge, from devices, from an architecture we haven’t even named yet? If so, today’s celebration could become tomorrow’s cautionary tale.