Swarm Current

Researching emerging technologies, overlooked companies & the infrastructure behind major shifts.

Sharing what catches my attention. Just curious — NFA.

[Swarm Current @swarmcurrent]

Swarm Current was started to follow developments that tend to get overlooked until they become obvious.

Right now, that means a lot of AI infrastructure, energy, and compute.

Mostly sharing what I'm researching, the companies catching my attention, and the questions I'm trying to answer.

[Swarm Current @swarmcurrent]

@StockSavvyShay Curious how much that compute will cost by the time it's actually delivered. A lot can change in GPU pricing and performance over the next few years.

Curious how much of that is already baked into these agreements 🧐

[Swarm Current @swarmcurrent]

Financing is another piece of the puzzle.

$CRWV bonds originally issued near a 9.7% yield reportedly traded around 15% last week.

Higher borrowing costs could make already-delayed projects even harder to finance.

This week's watchlist: grid commitments, production qualifications, and AI infrastructure debt spreads.

[Swarm Current @swarmcurrent]

WEEKLY PULSE ⚡

Last week's AI infrastructure takeaway: power delivery timelines are getting harder to ignore.

$AEP's 702MW Google fuel-cell project targets full operation in late 2028.

Microsoft exited Nscale's Loughton project, where grid power may not arrive until the 2030s. Its reason for leaving remains unconfirmed.

[Swarm Current @swarmcurrent]

The equipment side is moving, too.

Hitachi Energy committed C$479M to a Canadian transformer-insulation factory, targeting 2028.

Meanwhile, $NVTS unveiled a 24kW board converting 800V directly to 12V, targeting 97% efficiency.

Both developments are worth watching as AI power infrastructure scales.

[Swarm Current @swarmcurrent]

PULSE: In 2025, $Nscale announced a 50MW AI campus for Microsoft, targeting Q1 2027 GPU delivery.

Today, Microsoft reportedly exited the deal. Grid power may not arrive until the 2030s.

Contracted AI capacity deserves a closer look at when the power actually arrives.

[Swarm Current @swarmcurrent]

$NVTS and NVIDIA have developed a 24kW power board that converts 800V directly to 12V, eliminating the intermediate 50V conversion stage.

Target efficiency: 97% at full load.

Worth watching whether this architecture makes it into production systems.

[Swarm Current @swarmcurrent]

PULSE: The gap between contracted power and operational power deserves more attention.

$AEP is planning a ~$4B investment in 702 MW of Bloom fuel cells, backed by a 20-year Google power agreement.

Full operation isn't expected until late 2028 — that's a long gap between commitment and delivery.

[Swarm Current @swarmcurrent]

This could make already-contracted data center capacity even more valuable.

For example, $AIB. It secured a 15-year power agreement and recently signed a 50 MW deal with Nebius. As more projects run into permitting and political opposition, having power and customers lined up may become a serious competitive advantage.

[Swarm Current @swarmcurrent]

The financing structure is what interests me most here. CoreWeave is trying to bring insurance capital into GPU-backed lending, but its unsecured bonds are yielding around 15%.

Do you think lenders are questioning GPU collateral values, or the economics of the contracts supporting that debt?

[Swarm Current @swarmcurrent]

$NVDA contributed Vera Rubin rack and compute tray designs to OCP this week.

Meanwhile, Rittal and Bosch are demonstrating an 800V DC power sidecar designed to scale to 2MW.

Fascinating how quickly the hardware is evolving. Watching which designs actually reach production.

[Swarm Current @swarmcurrent]

$NVTS added a 24kW 800V→12V power board developed with NVIDIA, targeting 97% efficiency. It joins its earlier 800V→6V design.

Interesting that both voltage paths are being explored. The next signal I want is a production design win, not another demo.

[Swarm Current @swarmcurrent]

I keep coming back to is AI compute.

$CRWV has already shown how GPUs can be used as collateral for billions in financing. Now projects like GAIB and Aethir are experimenting with tokenizing GPU-related revenue.

Curious how much of the next wave of tokenized assets ends up being tied to physical AI infrastructure.

[Graham Stephan @GrahamStephan]

Honest question: how does the average person capitalize on this?

[Swarm Current @swarmcurrent]

@GrahamStephan One area I'd watch is AI compute. GPUs and data centers are expensive assets, and there's growing interest in financing them in new ways.

If tokenization makes it easier to finance or trade claims on that infrastructure, that could get interesting pretty quickly.

[Swarm Current @swarmcurrent]

Financing + what's next:

Capital is the other constraint: reports say SpaceX is seeking ~$40B for Nvidia GPUs. That's reported financing, not a closed deal. This week: $ASML results Wednesday and $TSM Thursday. Watch packaging capacity, lead times and capex discipline.

[Swarm Current @swarmcurrent]

The Undercurrent #001: Power, Packaging & the AI Buildout

Power:
$BKH laid out a 2.7 GW resource mix (including reserves) for Google's Cheyenne site.

$CEG and Google agreed to add 890 MW through nuclear plant upgrades. More AI power projects now spell out who builds, funds and manages the megawatts.

[Swarm Current @swarmcurrent]

Semis: $GFS signed a $2B deal to make silicon interposers for TSMC's CoWoS packaging. Volume ramp isn't expected until H1 2028. It's a real supply-chain shift, but too early to count as delivered capacity.

[Swarm Current @swarmcurrent]

@ZarinTasnifo3 👋

[Swarm Current @swarmcurrent]

One AI infra angle I think is getting missed: the old oilfield service companies.

$PUMP built distributed power for the oil patch. Now that same capability is being pushed toward data centers.

PROPWR already has ~510 MW contracted.

Worth watching where the next MW lands.

[Swarm Current @swarmcurrent]

Didn't realize $GFS was getting this involved in AI chip packaging.

They just landed a $2B agreement to manufacture silicon interposers for $TSMC

Production won't ramp until 2028, but interesting to see where the money is going beyond the actual chips.

https://t.co/CtoFQkezMr

[Swarm Current @swarmcurrent]

@mozo2say Agreed. That's what makes behind-the-meter power interesting. Could bypass some grid bottlenecks. Do you see that becoming a bigger advantage as demand grows?