Nearly Right

Nvidia pays $20 billion for Groq without buying Groq

How Silicon Valley learned to absorb competitors while regulators watch

"While we are adding talented employees to our ranks and licensing Groq's IP, we are not acquiring Groq as a company."

Read that again. Jensen Huang's Christmas Eve statement to Nvidia employees is a masterclass in legal precision. He acknowledges taking Groq's intellectual property. He confirms hiring its leadership. Then he insists none of this constitutes an acquisition. The contradiction is the point.

Three months earlier, Groq had raised $750 million at a $6.9 billion valuation. Nvidia paid $20 billion—nearly triple that figure. The premium wasn't irrational exuberance. It was the price of speed. A traditional acquisition triggers the Hart-Scott-Rodino Act: mandatory disclosure to federal regulators, thirty days of initial review, potentially eighteen months of investigation. Huang's carefully worded deal closed in days.

What Nvidia obtained was everything that mattered. All patents. All intellectual property. A perpetual licence to Groq's inference technology. Jonathan Ross, who invented Google's Tensor Processing Unit before founding Groq. President Sunny Madra. Roughly eighty per cent of the engineering team.

What they left behind was a corporate shell.

The playbook

Nvidia didn't invent this manoeuvre. They scaled it. Since March 2024, the same structure has appeared across Silicon Valley with the regularity of a franchise operation.

Microsoft went first, paying Inflection AI $650 million to "license" its technology while hiring co-founders Mustafa Suleyman and Karén Simonyan along with most of the startup's seventy employees. Inflection, once valued at $4 billion and training frontier models, became an enterprise licensing company overnight. The FTC opened an investigation. Germany's cartel office ruled it constituted a merger under their law. Neither jurisdiction blocked the deal.

Amazon followed in June 2024, absorbing two-thirds of Adept AI's workforce and its CEO while paying $330 million in licensing fees. Adept had raised $400 million to build AI agents. Today, four people list it as their employer on LinkedIn.

Google proved hungriest. In August 2024: $2.7 billion to hire Character.AI's founders—former Google employees who helped create transformer architecture—while licensing the startup's models. Then in July 2025, on the exact day OpenAI's exclusivity period to acquire Windsurf expired, Google paid $2.4 billion to poach the company's leadership and licence its coding technology.

Industry insiders now have a name for this: the HALO effect. Hire And License Out. Capture talent and technology. Avoid the complications of corporate absorption. Sidestep the regulators designed to prevent exactly this kind of consolidation.

The economics are brutally simple. The researchers capable of building frontier AI models number in the hundreds globally. Traditional acquisitions take years and face uncertain outcomes. This way, the deal closes before anyone can object.

The two-tier exit

When Google announced its Windsurf deal, approximately forty employees joined the search giant with substantial packages. Roughly two hundred others woke up to discover their company had been hollowed out while they slept.

Their equity—years of below-market salaries converted into ownership stakes, the startup dream made paper—was worthless. The corporate shell remained. The value had walked out the door.

One early Windsurf employee reported receiving one per cent of what their shares had been worth. Not one per cent less. One per cent total. Vinod Khosla, who has funded more successful startups than most people can name, called out the founders for "leaving their teams behind." He promised never to work with them again.

The outcry forced an unusual rescue. Within seventy-two hours, Cognition AI acquired Windsurf's remaining assets and employees, waiving vesting cliffs and ensuring everyone participated financially. "There's only one boat and we're all in it together," CEO Scott Wu wrote. The rescue made headlines precisely because it violated the new normal.

The Groq deal's employee outcomes remain murky. Some insiders claim all employees were cashed out fairly. Others question whether the $20 billion headline figure includes executive retention packages that would shrink the pool available for common shareholders. The uncertainty itself tells you something. In traditional acquisitions, the distribution of proceeds follows established rules. In these licensing-and-hiring arrangements, the rules are whatever the dealmakers decide.

What's clear is the emerging pattern. Founders and early investors receive full value, often at substantial premiums. Chamath Palihapitiya's Social Capital led Groq's 2017 seed round with $10 million at a $25 million pre-money valuation. Even after dilution across subsequent rounds, that stake is now worth between $1.6 billion and $2.4 billion.

The workers who built the technology occupy a different universe entirely.

The asset-strippers' playbook

This dynamic has a precedent, though not in technology.

When private equity firm Carlyle Group acquired HCR ManorCare in 2007, it didn't want a nursing home chain. It wanted the real estate underneath one. Carlyle loaded the company with debt, forced it to sell its properties, extracted value through management fees, then pushed the hollowed entity into bankruptcy—shedding pension obligations onto a government insurance programme while the firm had already recouped its investment.

The tactics differ in form. Private equity extracts physical assets: buildings, equipment, brands. Tech giants extract human and intellectual assets: researchers, patents, technology licences. But the underlying logic is identical. Separate what's valuable from what's burdensome. Extract the former. Leave someone else holding the latter.

Labour economists call this "fissuring"—the fragmentation of employment relationships into structures that preserve profit while externalising costs. The workers who built value find themselves in a corporate shell stripped of the assets that gave their ownership stakes meaning.

The Service Employees International Union documented this pattern across private equity: buyout deals generating "immense wealth for the private equity buyout industry" while creating "harsh consequences for workers." The critique applies with equal force to reverse acquihires. The mechanism is financial engineering. The outcome is wealth transfer from labour to capital, executed at the moment of maximum vulnerability.

The political web

The deal closed against a particular backdrop. David Sacks, Palihapitiya's co-host on the All-In podcast, became Trump's AI and Crypto Czar in December 2024. His financial disclosures reveal 708 technology investments, including 449 with AI connections. He serves as a "special government employee"—limited to 130 days of federal work annually, permitted to retain his portfolio.

In July 2025, Sacks co-authored "America's AI Action Plan," positioning artificial intelligence as a national security priority. Two months later, at the All-In Summit, Saudi Arabia's state-backed AI company presented Groq as exemplifying "the American AI stack in action."

The Saudi connection is itself notable. In February 2025, Saudi Arabia's Public Investment Fund committed $1.5 billion to expand Groq's data centre in Dammam. That infrastructure serves, in Groq's words, "nearly four billion people regionally adjacent to the KSA."

Nvidia's licensing structure excises this geopolitical complexity with surgical precision. By taking Groq's technology without buying the company, Nvidia avoids inheriting contracts that would trigger Committee on Foreign Investment review. GroqCloud continues independently, its Middle Eastern obligations remaining with the shell rather than transferring to America's dominant chipmaker.

The timing—Christmas Eve, when newsrooms run skeleton crews—ensured minimal immediate scrutiny of these connections. And 1789 Capital, where Donald Trump Jr. serves as partner, had invested in Groq's September round just three months earlier. That investment tripled overnight.

Coincidence or coordination, the pattern illustrates how concentrated AI dealmaking has become. A small network simultaneously shapes government policy, invests in AI companies, and profits from transactions that may benefit from relaxed regulatory enforcement.

Why the law cannot reach

The Hart-Scott-Rodino Act requires notification for acquisitions exceeding $126.4 million. Regulators then review whether the transaction would substantially lessen competition. The framework assumes acquisitions mean purchasing corporate entities.

Reverse acquihires exploit this assumption ruthlessly. No entity changes hands. The "non-exclusive licence" technically permits Groq to licence its technology to competitors—though without the engineering team that built it, such licences are nearly worthless. The hiring of employees, however strategic, isn't an acquisition under merger law. People can change jobs.

Germany's cartel office examined Microsoft's Inflection deal and concluded it constituted a merger under German law—then found they lacked jurisdiction because Inflection had insufficient domestic operations. The European Commission reached the same dead end. The UK's Competition and Markets Authority opened a probe that has blocked nothing.

The American situation is worse. The FTC investigation into Microsoft-Inflection continues without enforcement action. The expansive merger guidelines adopted in December 2023 face likely abandonment under the new administration. Whether reverse acquihires will receive meaningful scrutiny from regulators aligned with industry priorities is anyone's guess.

The fundamental problem is conceptual. Competition law addresses market consolidation through corporate combinations. When identical consolidation occurs through talent acquisition and technology licensing—when a dominant firm absorbs a competitor's capabilities without absorbing the competitor itself—the law's framework cannot grip the substance because it's designed to grip the form.

The inference war

Understanding Nvidia's willingness to pay such a premium requires understanding what's shifting in AI economics. Training large models—the phase where Nvidia dominates absolutely—demands massive parallel computation. Inference—running trained models to generate outputs—increasingly rewards different architectures: faster, cheaper, more energy-efficient.

Groq's Language Processing Units embody a fundamentally different approach. Rather than shuttling data to external memory, LPUs integrate large static RAM directly on the chip, enabling deterministic execution at dramatically higher speeds. Groq claimed throughput of 800 to 1,000 tokens per second for smaller models—many times faster than GPU inference.

Jonathan Ross created this architecture after designing Google's Tensor Processing Unit. He represents the vanishingly small pool of engineers capable of building inference-optimised silicon from scratch. By acquiring Ross and his team, Nvidia eliminates a competitor and blocks potential partnerships with Google, Amazon, or Microsoft—all of which are developing custom AI chips.

The $20 billion becomes comprehensible as insurance. Nvidia's data centre revenue exceeded $22 billion in a recent quarter. If inference workloads migrate to specialised architectures while Nvidia remains optimised for training, that dominance could fracture. Acquiring Groq's capabilities and the people who created them hedges against disruption. Eliminating a rival while doing so is a bonus.

The new normal

The numbers suggest this is neither aberration nor passing phase. Analysis tracking AI mergers recorded 177 deals in the second quarter of 2025 alone—roughly double the quarterly average since 2020. Startups acquiring other startups reached 427 transactions in the first half of 2025, up eighteen per cent year over year. Acquihires have become, in one analyst's phrase, "the defining deal shape" of the current era.

For employees, the implications are clear enough that industry observers now advise treating startup equity as worthless from day one. Demand market-rate salaries. Assume the upside won't materialise. Plan careers accordingly. The traditional bargain—sacrifice current income for future ownership—fails when deals are structured to pay insiders while leaving common shareholders in emptied shells.

For competition, the trajectory is equally stark. If dominant platforms can absorb competitors' talent and technology without regulatory review, consolidation proceeds regardless of antitrust law's intent. The gap between what the law addresses and what actually happens widens with each successful deployment of the reverse acquihire playbook.

Jensen Huang's phrasing—"we are not acquiring Groq as a company"—acknowledges what the deal accomplishes while insisting it isn't what it plainly is. That linguistic gap, between words and their meaning, between legal form and economic substance, is where $20 billion changed hands on Christmas Eve.

The regulators designed to prevent such consolidation watched it happen in real time.

#technology