High-VRAM GPUs defy the 2028 memory thaw; Meta pays Microsoft for models

Morning, humans.

Three stories today about who really sets the price.

Ordinary RAM should get cheaper once new fabs land after 2027. High-memory GPUs run on a separate clock, because the buyer is also paying for scarce packaging capacity, CUDA compatibility and deliberate product segmentation.

Meta spends hundreds of millions a year buying AI model access through Microsoft, while it trains its own models and weighs building a rival cloud service.

And California companies announced $366 billion in venture funding through August 20, nearly double the old record. Two companies supplied most of it. Seed funding went the other way.

Stay curious,

Marcus Schuler

A 96 GB RTX Pro 6000 sells for $16,000 against $1,999 for the 32 GB RTX 5090, and the extra memory inside it costs somewhere between $640 and $960.

New capacity lands in 2027, from Micron's first Idaho wafers to SK hynix M15X and Samsung's HBM4 ramp. Mainstream DDR5 is projected to run 20 to 45 percent below its cycle peak by 2029, with GDDR7 down 15 to 35 percent.

High-memory cards sit outside that cycle. Buyers pay for CUDA compatibility, scarce allocation, certification, and for one 96 GB pool that holds a job which would otherwise be split, shrunk or sent to the cloud.

Why This Matters:

• Anyone budgeting local inference hardware for 2028 should plan around VRAM that stays expensive while system memory gets cheaper.
• Segmentation rather than scarcity sets the top of the range, so a capacity glut will not close the gap between the tiers.

Meta buys hundreds of millions of dollars in annual AI model access through Microsoft Foundry, running trillions of tokens a week.

Neither company confirmed the spending, the token volume or the contract terms. Meta has used OpenAI models to grade its own systems' output while forecasting $130 billion to $145 billion in 2026 capital spending. The buyer of record here is also building the models it rents, and weighing an API service to sell them.

Answer.AI uses Peter Naur's theory-building account of programming to explain why coding models can shorten code without simplifying the system an engineer must understand.

A billing example supplies the missing context: a model recommended restoring Stripe subscriptions to fix automatic top-ups for Indian cards, while the team kept manual top-ups to preserve a payment system it could hold in its head. The piece turns code complexity from a metric into a product judgment.

• Broadcom is in talks with lenders to raise more than $60 billion in debt for an AI chip financing deal that would benefit Anthropic and other customers.
• Nevada cleared Tesla to deploy up to 5,000 robotaxis around Las Vegas over the next year, with Waymo and Uber capped at 1,000 each.
• Super Micro said an independent probe found no evidence its chief executive or senior managers knew of an alleged scheme to smuggle $2.5 billion of Nvidia chips to China.
• Apple Music told industry partners it will visibly label songs that content providers tag as materially generated using AI.
• Micro1 took its AI training-data business to a $500 million gross annual run rate from $100 million in eight months, with net run rate at $150 million to $200 million.
• GitHub blamed its seven-hour August 17 outage on a capacity failure after peak traffic overwhelmed an infrastructure component in a Central US data center.

Find the sales objection hiding in call notes. Teams often mistake a buyer's operational constraint for a product objection. This exercise separates the two before the next sales meeting.

Your raw input: three recent customer-call transcripts and a short description of the offer, price, and intended buyer.

The prompt:

Why this works: classification stops the model from treating every hesitation as resistance. Counting distinct customers limits one vocal buyer's influence, and quotations make the proposed test auditable.

What to use: Claude for long transcripts. ChatGPT works well for shorter calls.

Credit: Midjourney

Prompt: Cyberpunk aesthetic, a futuristic female android with her hair streaming in the wind, a striking pose in a cyberpunk city, mechanical, photorealistic, full body, cinematic lighting

OpenAI shipped an Apple Messages plugin for ChatGPT on macOS on August 20. It can read your chats, search them, analyze them, draft replies and send them. (9to5Mac, August 20, 2026)

Our take: The pitch is that you are too busy to answer your friends. The fix is a program that reads everything they have ever sent you and then answers in your voice. Somewhere on the other end a person is having a warm exchange with a statistical average of you, and thanking you for remembering their birthday.

The read-and-search half is the part worth pausing on. Your Messages history is the most complete archive of your private life sitting on any device you own, and it now has an API surface. Grant access once and the plugin keeps it. You get the convenience in six seconds, and the access stays granted until you go hunting for the toggle.

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai