torygreen

The energy demand for data centers is growing so fast that the grids can't keep up, and companies are now trucking in gas turbines to get around it.
APR Energy just rolled out eight mobile units, 1.1GW, aimed straight at data center demand.
Think about what that means. a turbine on a trailer is a worse deal than a grid connection in every way except one: you can have it running in months instead of years. paying that premium only makes sense if the wait is the thing killing you, and right now the wait is 5+ years in a lot of the US states.
So the buildout stops waiting for the grid and bring

Running DeepSeek V4 Flash 0731 on your own hardware is real now but what it costs and whether it's worth it depends a lot on who you are.
the machines that actually run it:
> 256GB M3 Ultra Mac Studio (~$7-9k): holds the 4-bit build with headroom, clean single box.
> 512GB M3 Ultra (~$10k+, if you can find one): runs the near-lossless version comfortably.
> 2x H200 or 4x A100 server: full checkpoint at 1M context, but now you're slightly in datacenter money.
it technically runs on a 128GB too, but only at heavy 3-bit compression that dents the quality.
so ~$7-10k for a serious single-box setup

Chip supplier. Infrastructure lender. Equity investor in its own customers. Nvidia is all three at once now, and this month the bill for holding all three hats hit $750B
> $250B backstopping OpenAI's Ohio lease.
> $500B entangled with SK Group on memory.
> $5B into SSI, a company with no product.
> $1B into Naver. A stake in Nebius. Chips flowing into Texas facilities it leases back to itself.
The loop Nvidia has essentially create is to fund the customer -> customer buys Nvidia chips -> Nvidia books the revenue -> Nvidia funds the next one. Supply and demand are collapsing onto the same ba

2026 is not the year closed labs planned for and the space where a closed frontier model can charge money is now about three points wide after K3 from @Kimi_Moonshot
K3 just landed cleanly above Opus 4.8 on capability, at half the price per task, with the weights going public in days. Only Fable 5 and GPT-5.6 Sol still sit above it. Everything below that line is now open territory, models anyone can serve at hardware cost, no R&D margin to recover.
That's the new math for every closed release. Land above the open frontier and you can price the gap. Land at or below it and you're charging a m

.@nvidia is putting 170,000 GPUs on an island in Indonesia. The campus was announced 2 weeks ago and goes live Q1 2027, a timeline thats practically impossible in US right now.
GPUs aren't the holdup, Nvidia is handing out 170,000 of them on a revenue share. What's scarce is somewhere to plug them in on a timeline that matters.
US interconnection queues run 5+ years in many regions. A behind-the-meter gas turbine takes 18 months. That gap is now the whole site-selection game.
And more friction keeps getting added on US side. Roughly $130B of capacity blocked or delayed in Q1 alone. Virginia,

$130B in data center projects blocked or delayed in Q1. That matches all of 2025 in just one quarter of this year.
75 projects. Opposition groups doubled from 396 to 833 across 49 states in three months. 300+ bills filed in the first six weeks of the year, moratorium proposals in 14 states, and New York's legislature just passed a one-year pause on anything over 20MW. Gallup has 71% of Americans opposing a data center near their home, up from roughly half a year ago.
Meanwhile hyperscalers plan to spend $690B on capex this year alone, as if none of this is happening.
Here's what makes this str

The most valuable real estate in AI right now is a shut‑down aluminum smelter in Kentucky.
@AnthropicAI signed a $19B, 20‑year lease for 401 megawatts at TeraWulf’s Hawesville campus, a former smelting site with roughly 480 megawatts of legacy grid connection. TeraWulf did about $1B in revenue last year; this single contract is worth almost twenty years of that.
The asset is the power hook‑up. You can manufacture more chips. You cannot create megawatts overnight.
Every bitcoin miner sitting on stranded power is turning into an AI landlord:
> Hut 8 has a $7B, 15‑year deal with Anthropic
> IREN

AI investment is outpacing AI revenue by 46%. During the 2001 telecom bust, that gap was 32%. We're past telecom-bust territory and still accelerating toward $1T in annual capex by 2027.
Token prices tell the other side. The Silicon Data LLM Token Expenditure Index peaked at $2.06 per million tokens in May. It's at $1.62 now. Down 20% in six weeks. Down over 90% since 2023.
Total token spend has roughly doubled year over year. Cheaper tokens are expanding the market, not destroying it. The problem is the capex bill doesn't shrink when prices fall. The infrastructure costs the same whether you

.@nvidia is no longer just in the chip-making game. It is positioning itself as a direct participant in the economics of AI infrastructure.
From the recent announcements, they're financing data centers and taking a share of the cloud revenue they generate.
Two recent partnerships illustrate the shift perfectly:
210,000 GPUs tied to $25–30B in committed offtake over six years
> Sharon AI $SHAZ deploying 40,000 GB300s in Australia
> Firmus scaling to 170,000 accelerators across 360MW in Batam
At the same time, Nvidia and partners introduced XFRA, the liquid-cooled GPU systems for distributed dep

70% of companies say they use AI. Fewer than 1 in 10 have an actual agent running in production.
That gap sits inside Stanford's AI Index, the most cited, least biased dataset on AI that exists, not published by a lab with a stake in the outcome.
Google alone spent over $150 billion on AI infrastructure last year. Frontier lab revenue is climbing at historic rates, and compute spend is climbing right alongside it, not shrinking as a share of revenue the way infrastructure normally does once something scales.
Adoption was never the bottleneck. The bottleneck is running a workload that never sto

The AI constraint in 2026 is not model quality. It is grid capacity.
PJM just printed an 833% jump in capacity prices for Virginia. That state now sends about 40% of its electricity to data centers and imports more power than California. Dublin is consuming close to 80% of Ireland’s national grid. Frankfurt is at 42% of regional supply.
The IEA thinks data center demand will roughly double again by 2030, while half of US projects under development are clustered in places that are already straining the grid.
We are getting better at squeezing joules per token and AI keeps getting more efficient

Spent time with the inference numbers and one stat keeps coming back.
70% of AI inference in 2026 runs at the edge. Not in a hyperscale data center. Not on AWS. On industrial systems, autonomous equipment, and connected devices where the round trip to a cloud server is not a latency problem, it is a failure mode.
The $106 billion inference market is not growing because people are sending more "chat" queries to their AI. It is growing because AI is being embedded into physical infrastructure that runs continuously, operates in environments with no reliable network, and cannot afford to wait.
Th

Apparently being close to the physical technology has very little to do with actually adopting it.
California is home to every frontier AI lab that matters. New York has more Fortune 500s than any other state. Both got lapped by Colorado, which hit 23.2% business AI adoption while New York managed 13.8%.
What Colorado and Arizona have isn't better infrastructure or talent. It's a willingness to move before the industry signals it's safe to. The states closest to the technology are often the slowest to deploy it because they have the most invested in how things already work.
Three quarters of A

On paper, AI costs should have collapsed by now.
The same frontier-level task that cost $30 per million tokens at GPT-4's launch in 2023 costs $1.25 today with GPT-5. A 96× drop in three years, across named models with published prices.
By any normal rule, enterprise AI bills should be cratering. Instead, average Fortune-500 AI budgets went from $7M in 2024 to $19M in 2026. Nearly 3× in two budget cycles.
The reason: cheap tokens don't get saved, they get weaponized. A chatbot fires tokens once and stops. An agent doing real work runs 5 to 30× more for the same task, and once inference is that

Something quietly inverted in AI compute this year, and it changes what the buildout is actually for.
In 2023, 2/3 of AI compute went to training, the actual work of building a model. The other, smaller slice went to inference, the work of actually running it once it's built. But that ratio quietly started flipping.
Inference is now 2/3 and still climbing, per Deloitte, and the chips built to run it crossed $50B this year.
The main reason this flip matters (and it's not percentage-wise): training and inference are different animals. Training happens in bursts, on one giant cluster, then it's d

Here's the split in AI compute that not many are reading correctly.
Frontier training is concentrating harder every quarter, thousands of GPUs that have to sit in one place wired together. But training is only 30% of demand in 2026. The other 70% is inference, and running it on a hyperscaler means paying for infrastructure built for the hardest workload to do the easiest one.
On distributed networks that same inference could run 45-75% cheaper and for anyone sizing an AI infrastructure budget, that gap is the whole story.
Training centralizes by necessity. Inference fragments because paying AW