The ad business as we know it is under serious threat and Amazon blocking Meta’s Muse is the perfect example of why.
If AI agents become the main interface people use to shop, search and make decisions, the entire advertising funnel starts to break. Instead of going to Amazon, searching for a product, scrolling through sponsored listings and clicking ads, you could simply tell an AI agent what you want and let it compare products, prices and reviews for you. At that point, Amazon risks becoming just another supplier while the AI agent owns the relationship with the customer.
Google could face the same problem on an even larger scale. Google makes money because people search for things and businesses pay to appear in front of those searches. But if people stop searching and start telling AI agents to find the best hotel, product, restaurant, flight or service for them, there are fewer searches, fewer clicks and fewer opportunities to show ads. The value starts shifting away from whoever owns the website or search engine and toward whoever owns the AI agent making the decision.
That is the bigger reason companies like Amazon have every incentive to fight this. AI agents do not just threaten traffic but rather threaten the entire business model built around controlling what users see before they make a purchase. If agents become the new front door to the internet, advertising as we know it could look completely different.
Hedge funds are back to shorting Software in size btw.
Today was the biggest increase in around a month for new Software shorts -- Muse clearly seems to be driving the market this week.
People are more scared of high friction / 0 moat companies that are subscription based.
Time will tell whether this is a trend especially as OpenAI drop their agent (this week?).....or just a one-off today.
I think it'll be the former.
Morgan Stanley’s Stephen Byrd full interview on CNBC: Power is the largest bottleneck, more than chips.
“Over the summer, the tech community reached out to all the U.S. time-to-power solution providers and said whatever power you have in ‘27 & ‘28, we will take it.”
$CIFR $WULF $HUT $GLXY $IREN $WYFI $NBIS
Morgan Stanley’s Stephen Byrd full interview on CNBC: Power is the largest bottleneck, more than chips.
“Over the summer, the tech community reached out to all the U.S. time-to-power solution providers and said whatever power you have in ‘27 & ‘28, we will take it.”
$CIFR $WULF $HUT $GLXY $IREN $WYFI $NBIS
A 'simplified' structure of Oracle's Project Jupiter SPV (4 data centers on 1400 acres in NM) providing $300BN in compute to OpenAi and funded by Bonds, Loans and Blue Owl equity.
The project is now hopelessly mothballed and may never see completion.
$NBIS GPU Rental Prices are up 20%
Let’s compare what these new On-Demand rates imply on a theoretical fully utilized $/MW basis to Dedicated Contracts (hyperscaler, enterprise, short term training).
🧧 What drives Pricing Premium?
Primarily by contract duration. The shorter the commitment, the higher the price per unit of compute, as $NBIS takes on more utilization and recontracting risk. Of course prepayments, credit backstops, and contract size also shape the economics.
That puts On-Demand at the top of the pricing stack. Instead of locking GPUs into dedicated contracts, $NBIS retains that risk and sells the same compute in smaller increments at materially higher unit economics.
◽️What GPU/Hour Pricing Actually Includes
These are standard AI Cloud GPU-instance rates, quoted per GPU/HR. Each instance comes with the GPU plus its allocated vCPU and RAM.
$NBIS also includes Managed Kubernetes, Soperator software, ingress/egress networking, and public IPs at no extra platform charge. Storage is separate, including block, object, and shared filesystem storage.
🏭 These are NOT Token Factory prices
@nebiustf is priced around managed inference consumption, which layers on incremental premium on top of base AI Cloud rates for model serving, model-specific tuning, and performance optimization
🎣 Caveats:
1. On-Demand rate assumes 100% GPU utilization
2. Dedicated contract rates are not directly comparable. As mentioned there are other factors beyond duration that influence pricing.
3. We also don't know silicon blend for these hyperscaler, enterprise, and short-term scale training contracts
4. B300 ≠ GB300 from a power-density standpoint. HGX B300 is ~1.1kW/GPU, while GB300 NVL72 runs materially higher. It’s unclear which configuration Nebius’ posted B300 rate refers to.
5. Posted On-Demand rates are list prices, not necessarily realized ASPs
6. On-Demand $/MW reflects GPU power only, not total facility power. CPU, networking, cooling, and PUE may not be subsumed, which would lower realized revenue per facility MW.