Petitions don't work. Except when they do.
Last year, a coalition of founders, lawyers, and startup associations launched EU-Inc.org - a grassroots push for a pan-European company structure. The petition gathered momentum. The policy proposal landed on desks in Brussels. And this week at Davos, Ursula von der Leyen announced the "28th regime" as an official Commission priority. And it's a big deal for European neoclouds.
Democracy:
Occasionally functional.
Here’s what’s inside this week:
Let’s get into it.
The GPU Audio Companion Issue #85
Want the GPU breakdown without the reading? The Audio Companion does it for you, but only if you’re subscribed. If you can’t see it below, click here to fix that.
Davos, hyperscalers and hegemons, and EU Inc.
Europe might finally have woken up to the realities of competition in 2026. And nowhere is that competitive pressure higher than in AI infrastructure.
European neoclouds currently face a patchwork problem. Expanding across France, Germany, and the Nordics means navigating different company structures, employment law, tax treatment, and capital regimes in each market. That means securing funding at a comparable scale to the US is challenging at best, and impossible at worst.
The solution?
The new regime, first proposed last year, includes significant reforms to capital markets and employee stock options. But that’s not the most interesting aspect. The regime also introduces a pan-European company structure, EU Inc., that would let businesses register, operate, and raise capital across all 27 EU member states under a single set of rules. The open question is execution, of course. But if EU Inc. ships, and the capital markets integration follows, a 450M-person market with unified structure starts to look competitive with the US in ways 27 fragments never could. Hyperscalers and hegemons might not be the only choices after all.
Why this matters:
European AI infrastructure companies and startups in general struggle to raise rounds of the scale that their US counterparts pull routinely.
The Savings and Investment Union, announced alongside EU Inc., aims to solve this problem by unlocking equity funding at a US-comparable scale.
All of this started with a petition and a massive, grassroots community effort from the team at EU–INC. The full industry proposal is worth reading if you're building, funding, or advising European AI infrastructure companies. It's available at EU-Inc.org.
Lightning AI and Voltage Park complete merger
One of the largest neoclouds by live capacity is moving up the stack. Fast.
Lightning AI and Voltage Park have merged, combining Lightning's MLOps software with Voltage Park's owned GPU infrastructure. The combined company has grown from $18M to over $500M in ARR since 2024. Lightning AI also built PyTorch Lightning, the deep learning framework used by 5 million+ developers. PyTorch Lightning wraps PyTorch to handle the engineering overhead (distributed training, checkpointing, logging, and hardware abstraction) so researchers can focus on model logic. It's become standard tooling for teams scaling from prototype to production. That developer base is now a funnel to owned infrastructure.
Why this matters:
Vertical integration is the playbook now. Owning both software and infrastructure creates a defensible margin. The hyperscalers have always understood this. Neoclouds are catching up.
Teams already using PyTorch Lightning can now deploy to Lightning infrastructure without switching tools, getting access to 35,000+ owned GPUs, including H100s, B200s, GB300s.
In turn, existing Voltage Park customers get access to new training, inference, and observability tools bundled in. Couple this with Voltage Park’s unique origin story, and you have the potential for some serious North American competition.
Sharon AI is reshaping its capital structure with three moves announced this week.
The Australian neocloud closed its $70M divestiture of a 50% stake in Texas Critical Data Centers LLC to New Era Energy & Digital. The deal, structured as a $50M senior-secured convertible note with $10M in cash and $10M in equity, ends Sharon AI's US footprint and positions the company as a pure-play Asia-Pacific compute provider - at least with current public info. At the same time, Sharon AI has also locked in up to $200M from Digital Alpha, a Cisco-aligned infrastructure investment firm, plus a three-way technology partnership with Cisco itself, and secured a $500M debt facility from USD.AI, a blockchain-native credit market for GPU-backed infrastructure.
Why this matters:
Rather than compete for hyperscaler overflow in Texas, Sharon AI could be betting on sovereign AI demand in Australia and APAC.
Combined with a $100M convertible note closed in December, Sharon AI has raised roughly $870M in committed capital over the past month. That's serious runway for an APAC-focused neocloud.
Throw in the Cisco partnership to dial up enterprise credibility, and a relative lack of regional competition, and you have a neocloud that may well find itself in exactly the right position in the coming months to take a sizeable share of the APAC market. Watch this space.
OpenAI wants a cut of your discoveries
OpenAI is done charging flat fees. Now it wants a share of whatever its models help create.
At Davos, CFO Sarah Friar declared 2026 the year of "practical adoption" and outlined a future where pricing is tied to outcomes, not just usage. The pitch: as AI moves into drug discovery, scientific research, and financial modelling, OpenAI wants pricing that "shares value with customers rather than just charging flat fees." Annualised revenue has already grown from $2B in 2023 to over $20B in 2025, tracking almost perfectly with compute expansion from 0.2 GW to 1.9 GW. But with over $1T in infrastructure commitments ($300B to Oracle, $90B to AMD), subscriptions alone won't cover the bill.
Why this matters:
Autonomous agents make outcome-based pricing possible. Unlike chatbots waiting for prompts, agents run continuously and execute across multiple tools. That's measurable work OpenAI can point to when asking for a percentage.
The enforcement question, however, is wide open. If a researcher brainstorms with ChatGPT, then runs experiments for two years, what share of the resulting patent belongs to OpenAI? Contract law isn't built for this.
This comes at the same time OpenAI is testing ads in ChatGPT for free users - something Altman previously called a "last resort."
Neurophos photonic chips promise 100x efficiency
Neurophos just raised $110M to replace your GPUs with light.
The Austin-based startup closed a Series A led by Gates Frontier, with M12 (Microsoft's venture fund), Aramco Ventures, Bosch Ventures, and Carbon Direct participating. The pitch: optical processing units that serve as drop-in GPU replacements for inference, claiming 100x the performance and energy efficiency of current silicon. The technical breakthrough is micron-scale metamaterial optical modulators, a 10,000x miniaturisation over previous photonic elements, letting them pack over one million optical processing elements onto a single chip. Funding goes toward datacenter-ready OPU modules, a full software stack, and early-access developer hardware.
Why this matters:
Photonic computing has been "five years away" for two decades.
The difference now: inference demand is hitting power walls that make exotic architectures economically viable. If 100x efficiency is real, data centre operators will find the budget for integration work.
Microsoft's Core AI Infrastructure VP called Neurophos "a breakthrough in compute on par with the leaps we've seen in AI models." When a hyperscaler's technical leadership endorses your physics, it's not just another alt-compute science project.
Nadella says the quiet part out loud
Microsoft's CEO just told Davos the AI boom needs to spread, or it's going to pop.
"For this not to be a bubble by definition, it requires that the benefits are much more evenly spread," Satya Nadella said during a conversation with BlackRock's Larry Fink. The "tell-tale sign" would be if only tech companies benefit rather than other sectors. Microsoft's own data backs him up. Their Global AI Adoption report (see Issue #83) shows productivity gains concentrated in wealthy economies. Nadella also reiterated Microsoft's multi-model strategy: after restructuring its OpenAI partnership in October and dropping exclusivity, Microsoft has been working with Anthropic, xAI, and others. His pitch to enterprises: the future isn't about picking one model provider.
Why this matters:
The infrastructure buildout assumes demand keeps scaling. If adoption stalls outside big tech and developed markets, data centre capacity under construction takes longer to fill. As we noted when covering the Bank of England's dot-com comparison (Issue #68), the maths only works if demand materialises.
Microsoft hedging its OpenAI bet has implications for its compute strategy. As exclusivity ends, Azure capacity could serve Anthropic, xAI, and others. The multi-model future Nadella describes is good news for neoclouds positioning as model-agnostic infrastructure.
The adoption gap isn't just a social issue. It's a market size constraint. If AI tools stay concentrated in wealthy economies, the total addressable market is smaller than the buildout assumes. That means oberbuilds, stranded capacity, and some likely heavy bags for investors.
Applied Digital breaks ground on 430 MW campus
Applied Digital is building first and signing the anchor tenant later.
The company broke ground on Delta Forge 1, a 430 MW AI Factory campus in an undisclosed southern US state. The specs: two 150 MW buildings supporting up to 300 MW of critical IT load across 500+ acres, with mid-2027 operations targeted. Applied Digital says it's in discussions with "another prospective investment-grade hyperscale customer" - corporate-speak for a major cloud or AI company that hasn't signed yet. Delta Forge 1 uses the same proprietary blueprint as their Polaris Forge campuses in North Dakota: repeatable execution, same design, different geography.
Why this matters:
Speed is the moat. As we covered in Issue #68, Applied Digital's $5B perpetual preferred equity facility with Macquarie provides it with capital to execute without the dilution that plagues smaller players. They keep winning by moving faster than most.
300 MW of critical IT load supports roughly 2,500 GB200 NVL72 racks at full density. That's enough for serious training clusters or large-scale inference deployments.
The location non-disclosure is strategic. Applied Digital is likely in final negotiations with both the hyperscaler and the state. We saw the same pattern with their $5B mystery customer at Polaris Forge 2 (Issue #70). Expect an announcement once the ink is dry.
The Rundown
Davos week usually produces more hot air than signal. This year was different.
Von der Leyen pitched EU Inc., a genuine attempt to fix Europe's fragmentation problem, born from a petition that actually worked. Nadella said what many have been thinking about bubble risk. On-stage with the CEO of BlackRock, no less. OpenAI telegraphed a revenue model that wants a cut of your discoveries, not just your subscription. And neoclouds kept consolidating: Lightning AI absorbed Voltage Park to build an integrated stack, Sharon AI exited the US with substantial fresh liquidity, and Applied Digital broke ground on another multi-hundred-megawatt campus before signing the tenant.
The theme?
Everyone's picking lanes.
Integrated stacks versus commodity compute. Outcome-based pricing versus subscriptions. Regional sovereignty versus global scale.
The buildout continues. But the strategic bets are narrowing. And the players who haven't picked a lane are running out of road.
See you next week.