Ready for 2026? We are. Big plans here for Epistrophy Capital. We expect a lot of the next twelve months to be focused on the dispersion of the AI buildout, new companies emerging in that arena and, yes, AI bubbles and frauds popping all around us.

The Consumer Electronics show will surely have lots of each (yeah, fake product demos are a big tell — we’ve seen it happen.) We also think the Nvidia (NVDA: NASDAQ)/Grok deal will show us just how weak the Federal Trade Commission has become under President Trump, as well as reveal some weaknesses in Nvidia’s product design, with big implications for Micron Technologies (MU: NYSE) — all themes of our report below.

NO MORE PASSWORDS! We’ve opened up the website with hopes that you’ll be able to search for the research you want more easily. Check it out: https://epistrophy.beehiiv.com.

As always, I’m focused on three things:
1) Technology-driven change;
2) the latest in innovation and startup trends, and;
3) stock fraud.

Companies Discussed

In This Note:

1. Companies Discussed

2. Red Zone: Nvidia Tackles Groq

   1. The Deal

   2. Yellow Flag: Under Further Review?

   3. A New Play Call? Nvidia’s Architecture Admission

   4. The Offensive Line: What This Means for Micron

3. Tweet O’ The Week

4. Epistrophy In The News

5. 📆 of Epistrophy Events

6. Availability This Week

Red Zone: Nvidia Tackles Groq

Will a replay review of the last-minute Nvidia’s Groq deal let it stand?

On Christmas Eve, as analysts and traders headed for the exits, news broke that Nvidia (NVDA: NASDAQ) had entered into a byzantine deal. The details are weird and there are some enormous implications to the announcement, not least: 

1. The likelihood of the deal being rejected by the Federal Trade Commission

2. Nvidia’s tacit admission that its architecture might be too much for AI inference;

3. A shift in the product lineup for memory makers like Micron Technologies (MU: NYSE). 

Forthwith: a look at all three. 

The Deal

Groq was seen by many as the biggest private challenger to Nvidia (our investment in the private Positron AI notwithstanding). So this deal was a surprise. A brief Groq press release described it as a “non-exclusive licensing agreement” covering Groq’s “inference technology.” Groq’s founder and chief executive Jonathan Ross, its president Sunny Madra and additional senior technical staff would join Nvidia. Groq stated that it would continue operating as an independent company under a new chief executive, Simon Edwards, and that its GroqCloud service would continue without interruption. Nvidia and Groq did not disclose financial terms or define the scope of the licensed technology beyond the phrase “inference technology.” 

Yellow Flag: Under Further Review?

The Federal Grade Commission might yet kibosh the entire deal. The transaction fits squarely within a category of arrangements that U.S. antitrust regulators have been scrutinizing since at least early 2024: licensing agreements paired with large-scale transfers of human capital that replicate many of the competitive effects of an acquisition without triggering formal merger review. In January 2025, the FTC released the (often heavily-redacted) results of a 6(b) inquiry into generative AI partnerships, explicitly highlighting deals that combine licensing, strategic payments and hiring as mechanisms that can “effectively swallow the start-up and its main assets without acquiring the firm.” That language is unusually blunt for an agency document and was clearly aimed at arrangements structured to avoid Hart–Scott–Rodino thresholds.

Nvidia is not unfamiliar to FTC enforcement actions. In December 2021, the FTC sued to block Nvidia’s proposed $40 billion acquisition of Arm, arguing that Nvidia’s control over Arm’s CPU architecture would allow it to disadvantage downstream competitors in multiple markets. The complaint emphasized Nvidia’s position as a platform company rather than a commodity chip supplier and framed the deal as an attempt to consolidate architectural control. Nvidia abandoned the transaction – paying a $1.25 billion fee to SoftBank – in February 2022 following opposition from regulators in the U.S., the U.K. and the European Union.

That history matters because the Groq transaction presents a cleaner fact pattern for a “quasi-merger” theory than Arm did. Arm was a horizontal acquisition with clear foreclosure risks. Groq is a nascent, inference-focused challenger whose relevance lies less in revenue than in architectural differentiation. By licensing Groq’s inference technology and hiring the people most closely associated with its design, Nvidia has acquired the capability to internalize Groq’s core ideas while materially weakening Groq’s ability to advance them independently.

Non-exclusivity does not neutralize that theory. Courts and regulators have repeatedly held that substance prevails over form when evaluating competitive effects. If the practical outcome of a non-exclusive license is that the incumbent gains broad rights to replicate the technology while the licensor loses the human capital necessary to compete, the agreement can still reduce future competition. The Groq deal fits squarely into the pattern regulators have examined but not yet unwound, similar arrangements involving Microsoft (MSFT: NASDAQ), Meta Platforms (META: NASDAQ) and Alphabet (GOOG: NASDAQ).

Nvidia’s market position raises the evidentiary stakes. We continue to talk to CEOs and other market participants and none have been willing to put Nvidia’s share below 95% of the global GPU market. When a firm with that degree of market power absorbs the core inference capability of a potential rival through a license-and-hire structure, regulators do not need to prove price effects today. They need only argue that a future competitive constraint has been neutralized.

The Groq transaction therefore sits in a legally exposed middle ground. It avoids the bright lines of acquisition review while triggering precisely the competitive concerns the FTC has been articulating in policy statements and staff reports. Whether that risk materializes into litigation will depend on internal documents, license scope and post-transaction behavior. Our prediction? The structure alone will not insulate Nvidia.

A New Play Call? Nvidia’s Architecture Admission

By licensing Groq’s inference technology, Nvidia acknowledges that inference bottlenecks now arise less from insufficient FLOPS and more from scheduler behavior, memory traffic and latency variance inherent in GPU-based systems. Is Nvidia’s GPU approach too much? 

Nvidia’s current inference platforms are built around high-bandwidth memory and system-level scale-up. The H200 GPU, introduced in late 2024, pairs the Hopper architecture with 141GB of HBM3e memory delivering approximately 4.8TB/s of bandwidth. Nvidia markets the H200 primarily as a memory upgrade over the H100, explicitly targeting large-model inference and long-context workloads. At the system level, Nvidia’s DGX B200 aggregates eight Blackwell GPUs into a single chassis with more than 1.4TB of total HBM3e memory and roughly 64TB/s of aggregate memory bandwidth. Nvidia’s rack-scale GB200 NVL72 platform extends this model by linking 72 GPUs in a single NVLink domain, which Nvidia describes as behaving like a single logical accelerator.

This design philosophy treats inference as a memory-dominated problem. As transformer models grow and key-value caches expand, performance is increasingly determined by how much model state can be kept resident in fast memory and how efficiently that memory can be accessed across devices. Nvidia’s response has been to increase HBM capacity per GPU generation and then mitigate the cost of distribution through faster interconnects and tighter system integration.

Groq’s architecture rejects that premise. Groq’s language processing units are designed around static scheduling and extensive use of on-chip SRAM as primary storage rather than cache. Instead of relying on dynamic hardware schedulers and runtime kernel dispatch, Groq’s compiler precomputes the execution graph, memory movement and inter-chip communication patterns ahead of runtime. Reuters characterized Groq’s approach as one that avoids external HBM, using SRAM to reduce latency and power consumption while limiting the size of models that can be served on a single chip.

The consequence is a fundamentally different trade space. SRAM is fast and predictable but expensive in area and limited in capacity. Groq systems therefore accept model partitioning as a first-order constraint and attempt to preserve performance through determinism rather than bandwidth. Static scheduling reduces tail latency by eliminating categories of runtime variability associated with thread divergence, cache contention and dynamic synchronization. The benefit is not higher peak throughput but tighter latency bounds, which directly affect fleet sizing and operating cost in production inference environments.

This distinction matters because inference economics are increasingly driven by tail latency rather than average throughput. A system that delivers slightly lower average tokens per second but predictable response times can be cheaper to operate at scale than a system with higher peak performance but greater variance. Nvidia’s GPU-based platforms attempt to suppress variance through scale-up and overprovisioning. Groq attempts to suppress variance through architectural constraint.

Nvidia’s decision to license Groq’s inference technology and hire its architects suggests Nvidia views this deterministic approach as complementary to, rather than substitutive of, its existing platforms. The public record does not indicate that Nvidia will manufacture Groq-derived silicon or replace GPUs for inference. What it does indicate is that Nvidia wants access to Groq’s scheduling, compilation and memory-management techniques as inference workloads become less forgiving of variability.

The Offensive Line: What This Means for Micron

The memory implications of the Groq transaction are indirect but significant, particularly for Micron Technology. 

A broad migration of inference away from high-bandwidth memory is unlikely, but that framing misses the more consequential shift now underway. The real change is not replacement but segmentation. Nvidia did not license Groq’s inference technology to hedge a speculative architecture. It did so because a growing class of inference workloads is constrained less by aggregate memory bandwidth than by execution predictability, scheduler overhead and latency variance—limitations that persist even in Nvidia’s most aggressive scale-up systems.

Nvidia’s current inference platforms attack these constraints through memory expansion and system aggregation. The H200 GPU, introduced in late 2024, pairs the Hopper architecture with 141GB of HBM3e delivering roughly 4.8TB/s of bandwidth and is explicitly marketed as a memory-forward upgrade over the H100 for large-model and long-context inference. At the system level, DGX B200 aggregates eight Blackwell GPUs into a single chassis with more than 1.4TB of total HBM3e memory and approximately 64TB/s of aggregate bandwidth. Nvidia’s GB200 NVL72 extends this approach to the rack scale, linking 72 Blackwell GPUs into a single NVLink domain that Nvidia describes as behaving like a single logical accelerator.

This design philosophy treats inference as a memory-dominated problem. As transformer models grow and key-value caches expand, performance increasingly depends on how much model state can be kept resident in fast memory and how efficiently that memory can be accessed across devices. Nvidia’s response has been consistent: increase HBM capacity per generation, then mitigate the cost of distribution through faster interconnects, larger NVLink domains and tighter system integration. Variability is suppressed through scale, redundancy and overprovisioning.

Groq’s architecture starts from a different premise. Groq’s Language Processing Units are designed around static scheduling and extensive use of on-chip SRAM as primary storage rather than cache. Instead of relying on dynamic hardware schedulers, speculative execution and runtime kernel dispatch—as Nvidia GPUs do through CUDA and TensorRT-LLM—Groq’s compiler precomputes the execution graph, memory movement and inter-chip communication paths ahead of runtime. Execution then follows that plan deterministically. Groq’s approach, however, explicitly minimizes reliance on external HBM, using SRAM to reduce latency and power consumption while constraining the size of models that can be served on a single device.

SRAM offers fast, predictable access – but is expensive and limited in capacity. Groq systems therefore accept model partitioning as a first-order constraint and attempt to preserve performance by eliminating runtime variability rather than by maximizing bandwidth. Static scheduling removes entire classes of latency variance associated with cache contention, thread divergence and dynamic synchronization. The payoff is not higher peak throughput, but tighter latency bounds.

That distinction now matters economically. Inference fleets are provisioned for tail latency, not benchmark averages. A system delivering slightly lower average tokens per second but predictable response times can be cheaper to operate at scale than one with higher peak performance and wider latency distributions. Nvidia’s GPU platforms suppress variance by adding HBM, NVLink bandwidth and hardware. Groq suppresses variance by constraining execution.

For Micron, HBM will likely remain indispensable for training workloads, frontier-scale models and long-context inference, all of which are anchored to Nvidia’s H200, DGX B200 and GB200 NVL72 platforms. There is no evidence Nvidia intends to reduce HBM content in those flagship systems. But at the margin—particularly in latency-critical, cost-sensitive inference deployments—the assumption that every incremental unit of inference requires proportionally more HBM is no longer unchallenged.

If Nvidia incorporates Groq-derived scheduling and memory-locality techniques into inference-specialized platforms alongside its GPU stack, HBM content per unit of deployed inference capacity could flatten even as total inference volume rises.

Nvidia is not questioning the value of HBM. It is questioning the universality of HBM-maximal architectures for inference. By internalizing Groq’s deterministic execution model, Nvidia is acknowledging that future inference scaling will depend as much on compiler control, memory locality and latency discipline as on raw bandwidth. For Micron, that redraws the boundary of where HBM remains irreplaceable—and where it may no longer be the default answer.

The Groq transaction is best understood as an architectural hedge rather than a conventional partnership. Inference has become the binding constraint in AI deployment, and deterministic execution and memory behavior have emerged as competitive variables alongside bandwidth and compute. Groq articulated a coherent alternative to the GPU-centric inference model by prioritizing predictability over flexibility. Nvidia’s response was not to compete with that approach in the market but to absorb it into its own ecosystem.

Whether that strategy survives regulatory scrutiny is an open question. The FTC has explicitly identified licensing-plus-hiring arrangements as a potential consolidation mechanism in AI markets, and Nvidia’s prior experience with Arm demonstrates that regulators are willing to challenge the company when it accumulates architectural leverage. The structure of the Groq deal appears designed to avoid formal merger review, but structure alone will not determine competitive effect. And Micron’s bet on GPUs might see a shift.

It’s clear that inference, not training, is now the name of the game. And Nvidia is taking an important competitor off the free agent market.

Tweet O’ The Week

Epistrophy In The News

On Schwab Network, I discussed how generative AI reshapes search economics and why Alphabet’s Google (GOOGL: NASDAQ) risks cannibalizing its own cash engine faster than Microsoft (MSFT: NASDAQ) does.

On Yahoo News, I broke down Nvidia’s (NVDA: NASDAQ) partnership with Groq, framing it not as a moonshot but as a $20 billion hedge against inference bottlenecks. We talked architecture, memory trade-offs and why Nvidia’s dominance forces it to experiment in public.

I also joined NewsNation on consecutive New Year’s Eves with guest host Laura Ingle to discuss where artificial intelligence actually goes next, beyond the slogans. The focus was on power, networks and capital intensity — the unglamorous constraints that will shape AI’s second act.

📆 of Epistrophy Events

Availability This Week

We’re back — and happy to talk about whatever “news” comes out of the Consumer Electronics Show. Feel free to text for time-sensitive items; email works too, replies later in the day.

Written reports are available to clients, with video summaries on Spotify and YouTube, a podcast on iTunes and of course our popular summaries of the summaries on Instagram, TikTok, and YouTube Shorts.