Nvidia's $20B antitrust loophole

Deal structure & regulatory arbitrage

• Nvidia didn’t buy Groq the company; it licensed Groq’s IP and hired key leadership/engineering talent.
• Commenters argue this structure likely avoids:
  • CFIUS review, given Groq’s large Saudi government contracts.
  • Formal antitrust merger review and its delays.
• Several see the ~$20B price (far above recent valuation) as paying for speed and certainty by dodging regulatory processes.
• Others note it’s not a “loophole” if regulators simply choose not to treat such acquihires as de facto mergers.

What Nvidia wants

• Groq is viewed as a serious inference competitor: an LPU architecture and chip reputedly better than GPUs for low-latency, high-throughput inference.
• Nvidia gains: IP, compiler stack knowledge, and the people who built and understand it.
• Some note that “non-exclusive” licensing may be largely cosmetic if Nvidia has all the top talent; others counter that IP can still be licensed to new implementers.

Fate of Groq, GroqCloud, and Saudi assets

• Many assume GroqCloud will be wound down over 12–18 months and the remaining company will wither.
• Others push back: Groq still has data centers, major Saudi commitments, and could survive as a cloud/infrastructure/IP-licensing business.
• Unclear from the thread whether Saudi investors are being cashed out and how much value remains in the “stub” company.

Employee equity, cap tables, and fairness

• A major thread: this structure may leave non-executive employees with worthless common stock while investors and founders capture most of the upside via secondary share sales or bespoke arrangements.
• Some argue employees in late-stage startups rarely hold more than ~0.01% and often don’t participate in such transactions at all.
• Others believe the headline price is high enough that common shareholders likely get something, but concede deal engineering could route most value to preferred holders and leadership.
• General advice trend: treat startup equity as having near-zero expected value; negotiate cash, and assume you may be excluded from liquidity events unless explicitly protected.

Antitrust, IP law, and state power

• Many see this as a case study in weak US antitrust: a dominant player can effectively absorb a serious rival’s brain trust and tech without merger review.
• Some contend that robust antitrust would treat such “IP + talent” deals like acquisitions when they have similar competitive effects.
• Others argue regulating where people can work would be unacceptable; that points to a mismatch between traditional antitrust tools (focused on corporate control) and modern competition centered on talent and IP.
• A few blame broader IP and corporate law for enabling consolidation; others propose tax/UBI schemes to disincentivize value extraction from labor.

Shifting startup and AI acquisition norms

• Commenters link this to a pattern in AI: “non-acquisitions” (licensing + key hires) instead of traditional M&A, seen as:
  • Easier on regulators.
  • More targeted: big companies buy only elite researchers/architects, not whole orgs.
• This is viewed as corrosive to the startup bargain: rank-and-file workers take risk and lower pay but can be cut out of big outcomes.
• Some predict labor will adapt by demanding higher salary, bonuses, and severance instead of banking on options.

Technical side notes

• Debate over whether Groq’s LPU is practically inference-only due to limited on-chip SRAM vs GPU HBM needs for training.
• Some challenge specific technical claims in the article (e.g., model sizes and throughput numbers; energy savings from reduced data movement), and a few suspect parts were AI-written or outdated.