Google’s $GOOG performance-per-dollar improvements with each TPU generation is helping it offer among the lowest “cost per intelligence task”, with Gemini 3.1 Pro Preview costing an average of $1.30, or 51% cheaper than GPT 5.5 (high) and 76% cheaper than Claude Sonnet 5.5. Google’s TPU 8i, delivering up to an 80% performance-per-dollar improvement over Ironwood, could extend this further. $NVDA $AMD

Google $GOOG: "Overall, our first-party model APIs are now processing nearly 32 billion tokens per minute, more than four times what we processed a year ago." This was also up 40% QoQ from 22 billion per minute in Q2. $AMZN $MSFT

Super Micro-backed neocloud Argentum AI is reportedly seeking a $5 billion credit facility to finance chip purchases, following competitor Nebius who secured $3.5 billion in financing last week. $SMCI $NBIS $CRWV

Google Cloud $GOOG is launching new Rubin-powered A5X instances, scaling to up to 80,000 $NVDA Rubin GPUs in a single site cluster and up to 960,000 Rubin GPUs in a multisite cluster, leveraging Google’s Virgo networking.

To defend gross margins, enterprise software platforms are now experimenting with new pricing models – with some like Oracle $ORCL absorbing the compute cost with free offerings and others like Salesforce $CRM charging per seat, per usage or per outcome. $TEAM $NOW

Groq is said to be considering using 3D stacked DRAM in next-gen chip architectures, in a possible shift away from its current SRAM-based designs as SRAM capacity is reportedly increasing just 10-20% every two years. $NVDA $MU $AMD

Google’s $GOOG merchant TPU sales are expected to follow a rapid growth trajectory through 2028, though analysts remain split on just how large this new revenue stream will be. Barclays forecasts a sharp ramp from $20 billion in 2026 to $252.7 billion by 2028, while JP Morgan and Piper Sandler see TPU revenue only just surpassing $100 billion by then. $NVDA $AVGO

Oracle $ORCL is also deploying Google’s $GOOG TPUs in its data centers for the first time, while also diversifying its inference-optimized compute with chips from Cerebras and Intel-backed SambaNova. $INTC $CBRS $NVDA

Meta's $META $25 billion stock sale is said to have gone five times oversubscribed, though the majority of shares went to existing investors with the top 15 largest stakes accounting for 70% of the offering, per Bloomberg. $NVDA $AMD $AVGO

Server CPU shipments are forecast to increase 31% YoY in 2026, accelerating to more than 40% growth in 2027 fueled by growth in inference workloads and increased agentic AI adoption, per TrendForce. $NVDA $AMD $INTC $ARM

Our firm’s back of the napkin math shows that Nvidia’s $NVDA ACIE segment could potentially overtake hyperscalers in two to six quarters. This is an important development as it results in Nvidia’s revenue being more resilient as hyperscalers shift capex toward designing and deploying their own custom chips. ACIE revenue grew faster than hyperscaler revenue last quarter at 138.5% compared to 101.2%, respectively. After reclassifying one company, the QoQ growth for ACIE was 25% compared to hyperscalers at 13%. Depending on how you model the two-quarter lumpiness, ACIE can conservatively overtake hyperscalers in 5-6 quarters at a 14% QoQ growth rate versus 10% QoQ growth rate for hyperscalers, or as soon as 2 quarters if the Q2 pace continues. In early September, Dell $DELL was in agreement with Nvidia’s results, stating that enterprises and neoclouds could contribute half of AI data center capacity growth through 2030, creating a $1 trillion total addressable market (TAM): “We’re expecting AI to be 75% of all data center demand by 2030, adding 200 gigawatts of power over that same time frame. And half of that, we believe, is right in our sweet spot with our customers, the Neoclouds, sovereigns and enterprises. And if you look at that math, we think the opportunity in front of us is more than $1 trillion over that time frame.” Although it may seem counterintuitive for Nvidia, Dell or Super Micro $SMCI to emphasize enterprise and neocloud customers over the hyperscalers and their $1 trillion capex next year, the reason is simple. Unlike the hyperscalers, these customers do not have those budgets to splurge tens of billions on custom chips. Instead, demand is more concentrated in ready-to-deploy servers that can immediately be brought online to generate revenue for immediate ROI, which Dell and Super Micro can do in as little as six hours to one day.

Meta's $META $25 billion stock sale is said to have gone five times oversubscribed, though the majority of shares went to existing investors with the top 15 largest stakes accounting for 70% of the offering, per Bloomberg. $NVDA $AMD $AVGO

Nvidia $NVDA and Oracle $ORCL are partnering to expand Vera Rubin NVL72 clusters on OCI to “multi-gigawatt scale”, with Oracle also introducing HGX Rubin NVL8 offerings and separate Vera CPU instances for agentic AI.

Oracle $ORCL is also deploying Google’s $GOOG TPUs in its data centers for the first time, while also diversifying its inference-optimized compute with chips from Cerebras and Intel-backed SambaNova. $INTC $CBRS $NVDA

Google’s $GOOG merchant TPU sales are expected to follow a rapid growth trajectory through 2028, though analysts remain split on just how large this new revenue stream will be. Barclays forecasts a sharp ramp from $20 billion in 2026 to $252.7 billion by 2028, while JP Morgan and Piper Sandler see TPU revenue only just surpassing $100 billion by then. $NVDA $AVGO

Morgan Stanley is projecting $AMZN AWS revenue to nearly triple from 2025 to 2028, with growth accelerating to 47% YoY in 2027 and remaining at 42% YoY in 2028, reaching revenue of nearly $374 billion versus $128.7 billion in 2025. $GOOG $MSFT

Despite consuming 2.6X more power per rack than the Maia 200, Microsoft’s $MSFT new Maia will pack 2X more chips per rack, deliver 2.5X more memory bandwidth and offer up to 7X higher throughput than the older generation. $GOOG $AMZN $NVDA