September arrives with conference season in full swing. Broadcom (AVGO: NASDAQ) and Zscaler (ZS: NASDAQ) report earnings this week, offering early signs of where enterprise spending may head into the fall. Last week was about Nvidia (NVDA: NASDAQ), a company now central not just to markets but to geopolitics. This week, attention turns to whether the broader ecosystem—semiconductors, cloud security, infrastructure—keeps pace with Nvidia’s momentum.

Have you check out our cool website? Visit the growing repository of past notes and our searchable research database at epistrophy.beehiiv.com.

As always, I’m focused on three things:
1) Technology-driven change;
2) the latest in innovation and startup trends, and;
3) stock fraud.

Companies Discussed

In This Note:

• Companies Discussed

• NVIDIA’s Ferrari Problem

  • Does everyone need a sports car in the AI garage? 

  • Training vs. Inference — Follow the Money

  • The Ferrari vs. the Workhorse

  • Inference Hardware: Ferrari vs. Workhorse

  • Why NVIDIA Sounds Defensive

• Tweet O’ The Week

• Epistrophy In The News

• 📆 of Epistrophy Events

• Availability This Week

NVIDIA’s Ferrari Problem

Does everyone need a sports car in the AI garage? 

NVIDIA’s earnings this week were another spectacle. The company posted record revenue and raised full-year guidance—but what really stood out in Wednesday’s Q2 FY2026 call wasn’t the topline growth; it was how much of the section on inference sounded like defense.

CEO Jensen Huang said, “Blackwell’s rack-scale NVLink and CUDA full-stack architecture address this by redefining the economics of inference.” For clarity: Blackwell NVL72 is NVIDIA’s rack-scale domain that connects 72 Blackwell GPUs and 36 Grace CPUs via NVLink, forming a tightly coupled, high-bandwidth “super-GPU.” CFO Colette Kress rounded out the pitch: invest three million dollars in one GB200 rack, she claimed, and you could generate thirty million dollars in token revenue—a tenfold return.

Again: nobody asked whether Blackwell was overbuilt for inference. NVIDIA seemed to preempt the question.

Training vs. Inference — Follow the Money

This matters because inference already leads the compute and energy budgets—and the gap is growing.

• Google reports that roughly 60 percent of its ML energy use is inference, not training.

• AWS estimates that 80 to 90 percent of its cloud ML demand is inference.

• Meta’s internal footprint shows a similar split.

Meanwhile, the Stanford AI Index Report 2025 (release in April)  confirms the cost to query a GPT-3.5-class model dropped more than 280-fold between 2022 and 2024, pushing consumption even higher. In this week's conference call, Huang himself acknowledged that “AI inference token generation has surged tenfold in just one year, and as AI agents become mainstream, the demand for AI computing will accelerate.”

Inference is not just the tail anymore—it’s the dog.

The Ferrari vs. the Workhorse

Blackwell NVL72 is a liquid-cooled, rack-scale system containing 72 NVIDIA Blackwell GPUs and 36 NVIDIA Grace CPUs, all interconnected by the largest NVLink network to date to act as a single, giant GPU for extreme AI and high-performance computing workloads. It’s an engineering beast. That is Ferrari-class compute, brilliant for training and extreme-scale inference runs.

But most inference workloads are bursty, millisecond-latency searches—far more amenable to workhorses: smaller, cheaper accelerators running quantized models.

Rivals are noticing. AMD’s MI300 (AMD:NYSE) and Intel’s Gaudi (INTC:NASDAQ) are positioned as inference-friendly alternatives. Hyperscalers are rolling their own: Google’s TPU v5e, tuned for large-scale inference inside Google Cloud (GOOG:NASDAQ), and Amazon’s Inferentia2, designed to lower the total cost of ownership for transformer workloads across AWS.

There are also private innovation plays. Positron AI (full disclosure: I’m an early investor in the company) claims its Atlas accelerator delivers 280 tokens per second per user for Llama 3.1 8B using just 2000 watts—versus NVIDIA’s 180 tokens per second at 5900 watts, making the Atlas roughly three times more power- and cost-efficient. Another is Groq, a well-funded startup whose Language Processing Unit chips are purpose-built for inference. They promise ultrafast, LLM-native performance with an inference-first ASIC design.

Inference Hardware: Ferrari vs. Workhorse

Why NVIDIA Sounds Defensive

That is why the call’s tone matters. Huang and Kress didn’t just market Blackwell—they defended it. They answered a question most investors haven’t fully asked: is Blackwell too much for inference? By reframing inference economics, they tried to make the Ferrari seem reasonable for everyday driving.

But the world is full of commutes—not track days.

Reasoning and agentic AI increase test-time token consumption, translating into more per-request compute—but that also intensifies the need for tokens per dollar efficiency, not just raw rack-scale horsepower. Cloud AI spending trends show inference budgets now exceed training by two to three times at major hyperscalers, with momentum only increasing in 2025. That is not a forecast; it is already visible in infrastructure accounting.

If the world bifurcates—Ferraris for frontier labs, workhorses for the rest—frontier systems like Blackwell will stay essential. But most customers will prioritize efficiency and economics.

Tweet O’ The Week

Epistrophy In The News

On NewsNation, I explained why Nvidia matters more than any other company right now—and why its relationship with China is at the center of U.S. technology policy. Washington wants to sell cutting-edge tools but not arm its rivals; Beijing wants access without dependence. Nvidia sits at the hinge of that uneasy trade-off, its chips both powering America’s AI boom and tempting adversaries.

📆 of Epistrophy Events

Availability This Week

I’ll be in San Francisco, attending Hubspot Analyst Day and meeting with companies and investors.

Written reports are available to clients, with video summaries on YouTube, and of course our popular summaries of the summaries on Instagram, TikTok, and YouTube Shorts.

Are these notes are helpful to you? Suggestions? I’d love to discuss them further and, as always, comments, questions and ideas are appreciated.