One-Line Summary

AI infrastructure money and strategy heat up: Firmus’ valuation jumps on Nvidia-fueled builds, Intel links with Musk’s Terafab while eyeing billion-dollar packaging wins, Meta inches toward a hybrid open strategy, and OpenAI both invests in multi-agent swarms and escalates its legal fight with Musk.

Big Tech

OpenAI Pushes AGs to Probe Musk Ahead of Trial

OpenAI, the lab behind ChatGPT, asks the California and Delaware attorneys general to investigate “improper and anti-competitive behavior” by Elon Musk and associates, weeks before jury selection begins on April 27 in the Northern District of California. The letter from OpenAI strategy chief Jason Kwon alleges coordinated attacks to undermine OpenAI, including with Meta’s Mark Zuckerberg. The stakes are high: Musk seeks damages exceeding $100 billion tied to OpenAI’s restructuring into a for-profit entity. ¹

OpenAI argues Musk’s actions could derail its mission to ensure artificial general intelligence (AGI) benefits humanity, warning that attacks aim to shift control of AGI from mission-led entities to competitors “who lack mission-driven principles.” In January, OpenAI told investors to expect “attention-grabbing claims” from Musk as trial nears, framing the dispute as both commercial rivalry and governance test for AI nonprofits turned hybrids. ²

A separate account notes OpenAI granted Microsoft a 27% stake as part of its restructuring and that state AGs reviewed governance commitments before not opposing the shift. The letter also claims Musk tried to seize control of the nonprofit and even solicited Zuckerberg in a failed takeover attempt, sharpening the narrative that legal and regulatory arenas are now central battlegrounds in the AI race. ³

Meta’s ‘Hybrid Superintelligence’: Fewer Fully Open Models?

Reports suggest Meta’s next-generation AI models, developed under Alexandr Wang (ex-Scale AI), could debut as closed models first, with open-source variants to follow—marking a shift from the broadly accessible Llama series toward a more “hybrid” stance that balances openness with safety and competitive performance. Axios coverage cited by Social Media Today says the new models aim to close benchmark gaps after Llama lagged rivals. ⁴

Why it matters: Meta has used “open” to win developers and distribution across Facebook, Instagram, and WhatsApp, but safety, IP, and product differentiation pressures are rising. Some outlets describe a pragmatic approach—initially closed to manage risk and protect specs, then selectively open—echoing an industry pattern where the most capable “frontier” models trend more proprietary. ⁵

The Superintelligence group reportedly nears first releases, positioning for consumer-facing strengths (e.g., shopping tools) while chasing performance parity. Skeptics note internal debates over paths to AGI and whether current generative approaches are sufficient. Still, if Meta ships a strong consumer AI layer soon, marketers and creators could see faster feature velocity across Meta’s apps. ⁶

Industry & Biz

Firmus, Nvidia-Backed ‘Southgate’ Builder, Hits $5.5B–$7B Valuation Firmus, an Asia-Pacific

AI data center developer, raises $505 million led by Coatue at a $5.5 billion valuation, lifting its total six-month haul to $1.35 billion. Nvidia participates as Firmus accelerates deployment of Nvidia’s forthcoming “Vera Rubin” AI factory design, aligning with CEO Jensen Huang’s push for “sovereign AI” data centers that keep national data onshore. ⁷

The flagship “Southgate” program starts in Tasmania, aiming to power facilities with renewables and, after the first two tech rollout waves, house computers built on about 36,000 Nvidia accelerator chips. That’s meaningful scale for model training and inference, and signals hyperscaler-grade customers are in the mix. The data center buildout could translate into lower latency and more regional AI capacity for enterprises in Australia and Singapore. ⁷

Valuation momentum is intense: TechCrunch pegs Firmus’ raise tally at $1.35 billion over six months, while separate reports say Nvidia is doubling its investment as Firmus closes a $1 billion-plus pre-IPO round, suggesting a valuation approach toward $7 billion—up from $1.9 billion in September 2025. A planned IPO pitch to Asian investors underscores how “AI real estate” is maturing into its own asset class. ^{8 9 10}

For users and teams, Firmus’ build aligns with tight global GPU supply: more local capacity can stabilize access, pricing, and compliance for regulated workloads. Strategically, Nvidia’s investment loop—backing companies that also buy its chips—invites debate about circularity, but it clearly accelerates infrastructure that keeps Nvidia’s ecosystem dominant. ⁷

Intel Joins Musk’s ‘Terafab’; Packaging Could Bring Billion-Dollar Wins

Intel says it will join Elon Musk’s Terafab mega chip complex with SpaceX and Tesla, aimed at processors for robotics (like the Optimus humanoid) and data centers—including Musk’s ambitious idea of orbital data centers to bypass Earth-based power and cooling limits. Reuters frames this as part of Musk’s vertical integration to control AI chip supply amid soaring demand. ¹¹

Beyond Terafab, Intel’s near-term cash cow may be advanced packaging—the back-end process that stitches chiplets and high-bandwidth memory into final AI packages. Wired and Yahoo Finance report Intel is in advanced talks with Google and Amazon, with CFO Dave Zinsner signaling potential “billions per year” packaging revenue at roughly 40% gross margins—arriving faster than wafer foundry deals. ^{12 13}

Why it matters: AI performance now hinges on packaging capacity (think CoWoS alternatives), not just cutting-edge nodes. If Intel closes hyperscaler packaging deals while collaborating on Terafab, it gains leverage as a U.S.-based alternative to TSMC bottlenecks—useful for companies seeking domestic supply resilience and export-control insulation. ¹³

OpenAI Backs Isara’s ‘Agent Swarms’; Mega-Fundraise Reports Emerge

Isara, a 9‑month‑old startup from ex-OpenAI and Oxford founders, raises $94 million at a $650 million valuation to coordinate thousands of AI agents for complex analytics like commodity forecasting. The vision: move beyond single-model prompts to “teams” of specialized agents that communicate, divide work, and converge on answers—initially targeting hedge funds, then biotech and geopolitics. The technical challenge is avoiding cascading errors and misaligned goals at swarm scale. ¹⁴

OpenAI’s participation is strategic optionality: if multi-agent architectures become essential, a minority stake keeps OpenAI close to outside breakthroughs and talent. It also fits the “neolab” trend—research-heavy startups raising at high valuations pre-revenue on the promise of novel architectures that could leapfrog today’s large language models. The next 12–18 months will test whether demos translate to dependable production systems. ¹⁴

Separately, trade outlets report OpenAI has secured an unprecedented $122 billion in committed capital at an $852 billion valuation, with anchors including Amazon, Nvidia, SoftBank, and continued Microsoft participation—alongside a larger undrawn credit facility and claims of $2 billion in monthly revenue. These figures, if borne out, suggest an aggressive infrastructure build (multi-cloud, multi-silicon) and an enterprise pivot that could reshape vendor choices across the stack. ^{15 16}

What This Means for You

• Capacity is strategy: More regional data centers (Firmus) and new packaging capacity (Intel) mean better odds you can secure GPUs and memory-bound AI capacity on timelines that match product roadmaps. If you’ve hit allocation walls, APAC options and U.S.-based packaging routes could reduce wait times and diversify risk. ^{7 13}

• Verticalization is back: Musk’s Terafab and Intel’s packaging push show how mission-critical AI is pulling compute closer to the product. Expect tighter integration between hardware, models, and apps—great for performance, but it can lock you into ecosystems. Negotiate exit ramps and data portability up front. ^{11 12}

• Openness is contextual: Meta experimenting with “closed-first, open-later” suggests enterprises may see faster features but less latitude with the cutting edge. Plan pilots with model-agnostic layers and keep an eye on license terms—today’s open can become tomorrow’s hybrid. ^{4 6}

• Agents are getting organized: If multi-agent “swarms” mature, your analytics, research, and ops could move from single prompts to orchestrated AI workflows. Start small with multi-agent frameworks, measure error propagation, and set guardrails before entrusting P&L-impacting calls. ¹⁴

Action Items

1. Map your 2026–27 GPU and HBM needs: Engage providers in APAC and U.S. packaging vendors to compare pricing/lead times; include Firmus-style regional data centers on your RFP list. ⁷
2. Prototype a multi-agent workflow: Use existing frameworks to orchestrate 3–10 agents on a real task (e.g., KPI forecasting), and log failure modes to inform whether “swarms” warrant budget next quarter. ¹⁴
3. Negotiate model-agnostic contracts: As Meta and others go hybrid, ensure your AI stack (vector DBs, orchestration, evals) is portable across at least two model vendors. ⁶
4. Explore Intel advanced packaging options: If you build custom silicon or rely on chiplets/HBM, book a discovery call on EMIB/Foveros timelines to hedge against CoWoS constraints. ¹³