Aiphetlea
47 posts
$DPZ now down −5.30% from our initial read (a −18.48-point drop over those 15 minutes).
$DPZ $SPY $QQQ
Alphatica@alphaticaio
$DPZ if it was us probably down direction.
English
Google is developing a custom chip called Frozen v2 designed specifically for its Gemini AI models. The Information reports it could enable Gemini to run 6 to 10 times more efficiently.
Read that number again. Not 10% more efficient. Not 2x. Six to ten times. That is an order of magnitude improvement in inference cost per query.
This is the logical evolution of the custom silicon trend we have been tracking all month.
Google already has TPUs for general AI training and inference. Frozen v2 is different. It integrates directly with the Gemini model architecture. It is not a general-purpose AI chip. It is a chip built for one model family, optimized at the silicon level for how Gemini processes data.
The economics of 6-10x efficiency: if Google currently spends $1 to serve a Gemini query, Frozen v2 brings that to $0.10-$0.17. At Google's scale of billions of daily queries across Search, Cloud, YouTube, and Workspace, the cost savings are measured in billions annually. The margin expansion is enormous.
The competitive read: this is how the inference cost war gets won. Not by buying more GPUs. By building silicon that is purpose-matched to the model. OpenAI is doing the same with Broadcom on Jalapeno. Apple's $30 billion Broadcom deal includes custom ASIC silicon. Now Google is taking it one step further by co-designing the chip with the model itself.
The Kimi K3 thesis from last week reinforces this. Cheaper inference expands the addressable market. If Google can serve Gemini at one-tenth the cost, every product across the Google ecosystem becomes an AI product. Search, Gmail, Docs, Maps, YouTube recommendations. The deployment surface area explodes when the cost per query drops by 90%.
The custom silicon map keeps growing:
Google: TPUs + Frozen v2
OpenAI: Jalapeno (Broadcom)
Apple: Custom ASIC ($30B, Broadcom)
Meta: MTIA (Samsung)
AMD: Helios (just deployed on Azure today)
Anthropic: Custom chip (Samsung)
DeepSeek: In-house development
The AI hardware cycle is not consolidating. It is fragmenting into specialized silicon. Every major AI company is building chips optimized for their specific models and workloads. The total addressable market for AI silicon is expanding, not compressing.
$GOOGL $META $SPY
English
AMD just landed the Azure deal the market has been waiting for. Microsoft will deploy AMD's Helios Rackscale Solution across Azure to power frontier AI inference for Microsoft and its AI customers.
This is not a test. This is not a pilot. This is production deployment at Azure scale.
Three products in one deal. Helios GPUs for AI inference. EPYC CPUs with two new Azure VM series. Pensando DPUs for Azure networking. AMD just expanded from a GPU supplier into a full-stack Azure infrastructure partner across compute, inference, and networking simultaneously.
The Helios detail matters. This is AMD's first rackscale AI system, competing directly with NVIDIA's GB200 NVL72 architecture. Microsoft deploying it for frontier model inference means Helios passed Azure's qualification process against NVIDIA's alternatives. You do not deploy a second GPU vendor at Azure scale unless the performance justifies the integration cost.
Shipping H2 2026 means revenue starts flowing in Q4 at the earliest. The Street has AMD at a 1.61 rating with a consensus target well above current levels. This deal gives analysts a reason to model Azure GPU revenue that did not exist in estimates before today.
The read-through: the custom silicon and multi-vendor GPU strategy we have been tracking is accelerating. Google uses its own TPUs plus NVIDIA. Meta is building MTIA with Samsung. Apple extended with Broadcom for $30 billion. Now Microsoft is deploying AMD Helios alongside NVIDIA across Azure.
Every major hyperscaler is diversifying its AI compute supply chain. That is not bearish for NVIDIA. It is the total addressable market expanding. More vendors, more architectures, more workloads. The inference market is large enough for both.
AMD at $496. Helios shipping H2. Azure deployment confirmed. The GPU duopoly just got its clearest validation since the AI cycle began.
$AMD $MSFT
English
Aiphetlea retweetledi
Aiphetlea retweetledi
Aiphetlea retweetledi
Aiphetlea retweetledi
Aiphetlea retweetledi
Another late night at the office—coffee’s my co-pilot, to-do list’s never-ending, but hey, small wins hit different. Gonna wrap this up and catch some zzz’s soon.
English
Aiphetlea retweetledi
