Everyone now wants a piece of all that cash the memory companies will make in the next coming years.
Around 2:00 PM Korea Standard Time (KST) on September 3, 2026, the Korea Electric Power Corporation (KEPCO) publicly issued a massive 25 trillion won ($17 billion) electricity prepayment proposal to SK Hynix and Samsung Electronics.
Debt-ridden state utility KEPCO requested that SK Hynix prepay 5 trillion won ($3.7 billion), while asking Samsung to prepay 20 trillion won.
The Timeline: The proposal asks both tech giants to pay for 5 years worth of electricity bills in advance (covering 2027 through 2031).The Strategic Objective: KEPCO plans to use this upfront cash influx to finance the immediate construction of massive power grids required for South Korea's expanding AI and semiconductor manufacturing clusters.
We will see a lot more of these types of events in the next coming years - but none of that will matter. The cash will keep flowing to SK Hynix/Samsung
Just signed up for GrokBot today - asked one of my agents to look for relevant tech news on reddit - gave me some subreddits that I think i should check out.
This is actually very good stuff.
Everyone talks about hyperscaler and neocloud backlog like it represents broad, diversified enterprise demand. Increasingly, it does not. A massive share of future AI infrastructure revenue is now tied to just two private frontier labs: OpenAI and Anthropic.
Across the top five neoclouds, total estimated backlog is around $275B to $300B+. Direct OpenAI and Anthropic exposure is roughly $115B to $128B, or 42% to 46%. Including hyperscaler pass-throughs, exposure rises to roughly 52% to 56%.
When a product that wasnt really selling out before suddenly gets used for things other than what it was designed for... that stock gets repriced eventually. Apple at below $5 trillion until now? Not for long.
I don't know but this supposed correction in the market kinda is not really panning out. Just look at SK Hynix Samsung and the Kospi intraday today - from down 4% to flat to end the day:
"OpenAI's Jalapeno chip beats Nvidia's Rubin chips on inference." was the headline everyone was focusing on last week.
OpenAI designed the chip with pre-GPT5 models. Can you imagine what the third Jalapeno chip will be like with GPT 5.6 Sol doing the design work for them now?
The irony of all of this - Nvidia chips trained the models that designed the chips that will replace them eventually. If that is not a classic case of "Recursive Self Improvement" then I don't know what is anymore.
Don't sleep on Apple. Everyone still clinging on to the narrative that they have missed the AI revolution- but their silicon combined with MacOS and iOs is the best for consumer edge.
They will give Nvidia a run for its money for first company to $10T in the next 5 years.
Nvidia buys Huggingface for $13B
So in a space of a week - Jensen bought Poolside (the factory for models), Perplexity (AI Agents and interface), and now Huggingface - where all these models are accessed by the market.
These three companies would have IPO'd by now if not for this trend of selling to the big tech companies. The market gets none of the upside - the only way to play them is via the Nvidias and the Googles etc.
Don't know how to feel about this TBH.
When you break it down to an hourly rate, Anthropic paid $7.79 per available GPU-hour on Elon Musk's Colossus cluster. By comparison, they are paying $5.84 per available GPU-hour on the Nscale deal.
If you look at the market rates below, it's pretty much at market price (except for the Colossus 1 deal which are mostly Hoppers I believe).
So the rate for compute that comes in about a year from now is at $16.3M per MW (Nscale). Whereas the spot rate is about $50M per MW (SpaceX).
Very interesting. Anthropic and OpenAI are ushering in the birth of a new asset class with its own spot and futures pricing mechanism. Pretty sure Chicago Mercantile Exchange will have this commodity (compute) trading in less than 6 months.
Anthropic is paying $45B over six years for 460MW of Vera Rubin compute at Nscale’s West Virginia campus.
Contract economics:
>$45B ÷ 6 years = $7.5B/year
>$7.5B ÷ 0.46GW = $16.3B/GW-year
>Nscale maps the initial 1.35GW to ~430,000 GPUs (PR):
430,000 ÷ 1,350MW = ~319 GPUs/MW
460MW × 319 = ~146,500 Anthropic GPUs
$45B ÷ (146,500 GPUs × 8,760 hours × 6 years) = $5.84 per available GPU-hour, or $6.87 at 85% utilization.
The initial phase is three buildings totaling 1.35GW of IT capacity. The first comes online in late 2027, with the others following in 2028.
Estimated capex:
>$47B GPUs (66%)
>$17.7B data centers
>$6.5B onsite power
= ~$71B total
That equals $52.6B/GW including GPUs, or $17.9B/GW for the physical infrastructure alone. Still hovering around that ~$50B capex per GW figure.
The campus is being energized by 2GW of onsite natural-gas generation plus batteries, designed to operate independently of the public grid.
Salesforce teams up with Anthropic to bring you "Claudeforce". So the one holding all your company data (CRM) will give it all to Anthropic so they can get the bets parts of it and figure out who is doing well so they can then launch a rival product.
What would have been a better product for everyone is Salesforce teams up with AWS, Azure, and GCP - then offer an opensource LLM product and then Zero Data Retention. I think this is a huge opening for a startup to start disrupting CRM.
I’m thinking about banning Claude code at Shopify until they change their mind and read AGENTS.md and .agents/skills etc.
Insisting on only reading CLAUDE.md sometimes leads to split brain problems when different team members use different tools. Just unnecessary.