AI in the news: week of December 28, 2025

What this week actually changed: Nvidia removed a credible inference competitor without an antitrust hook, and the open-coding gap closed enough to make hosted-frontier the exception rather than the default.

Christmas week. The news cycle takes a half-step back and most of what runs is either pre-cooked year-in-review content or stories held until the holiday lull cleared. That said, this week produced one genuinely large piece of business news (the Nvidia-Groq deal) and one model release that matters more than its quiet drop suggests, which is GLM-4.7 from Z.ai. Shorter roundup, partial year-in-review setup. The formal 2025 wrap-up sits at 2025 wrap-up: what actually changed in AI, and next Sunday will be the first roundup of 2026 with a proper look-ahead. This week is the bridge.

On December 24, Groq announced a non-exclusive licensing agreement with Nvidia for Groq's inference technology. The structure is technically a license-plus-asset-purchase rather than an outright acquisition. Groq the company continues to exist, GroqCloud isn't part of the deal, and finance chief Simon Edwards becomes CEO. But the substance, per CNBC reporting, is Nvidia paying roughly $20 billion in cash for Groq's assets and hiring founder Jonathan Ross, president Sunny Madra, and the rest of the senior technical leadership. This is, by a wide margin, the largest deal in Nvidia's history. The prior record was the ~$7B Mellanox acquisition in 2019. The structuring is the interesting part. As one analyst put it to CNBC, the non-exclusive-license-plus-leadership-hire shape is built to "keep the fiction of competition alive." Read that as: Nvidia is removing one of the few credible non-GPU inference competitors from the market without triggering the antitrust review an outright acquisition would invite. The Motley Fool's "aqui-hire" framing is the right read.

I want to be clear about why this matters. The thesis I've been pushing for fifteen months is that distributed beats concentrated, and that the structural risks of frontier AI come less from the models and more from the consolidation of the underlying infrastructure. The Nvidia-Groq deal is the consolidation thesis playing out in plain view. Groq's LPU architecture was the most credible bet on inference being a fundamentally different problem than training, with a fundamentally different chip optimized for it. That bet now lives inside Nvidia, structured as licensing rather than ownership specifically so the regulators don't have an easy intervention hook. The longer-term consequence: inference is where the cost curve will live for the next five years. Training costs are large but bounded; inference costs scale with usage, and usage is exploding. If the inference-chip market consolidates into Nvidia at the same rate the training-chip market did, the bargaining position of every AI-using business gets materially worse. The on-prem and open-model story I keep coming back to in the on-prem case and the vendor-lock-in piece gets harder to execute when the alternatives quietly absorb into the incumbent. Groq was an alternative. As of December 24, it's a part of Nvidia. I don't think the deal gets blocked. The license-not-acquisition framing is too clean and the current enforcement posture too light. I do think we'll look back on this week as one of the consolidation milestones of the AI era.

December 22, the other story. Z.ai released GLM-4.7, an open-source large language model targeted at production coding. Top-ranked open model on Code Arena, top open-model result on τ²-Bench (87.4 for interactive tool use), and, per GIGAZINE's coverage, beats Gemini 3 Pro on a number of coding benchmarks. The weights are on Hugging Face. Z.ai is also pricing the hosted version at $3/month and announced intent to be the first publicly listed AGI foundation-model company on the Hong Kong stock exchange. The benchmark numbers are what they are, every model release comes with cherry-picked wins, and GLM-4.7 won't actually beat GPT-5.2 or Sonnet 4.5 across the board. What matters about this release is the gap. Twelve months ago, the conventional wisdom was that the best open coding model was meaningfully behind the best closed one, and the gap was widening. Twelve months on, the gap has narrowed to "competitive on most benchmarks, ahead on some: $3 vs. $200." DeepSeek R1 in January was the canary; the trajectory through Qwen, Mistral, and now GLM-4.7 is the trend.

The principled-practitioner read is the same one I've been giving since spring: if you're standing up an internal coding assistant in 2026, you should be running GLM-4.7 or a comparable open model on your own infrastructure as the default and reaching for the hosted frontier labs only for the queries the open model can't handle. The cost math, the data-residency math, and the lock-in math all point the same direction. The hosted-frontier model is the exception, not the rule. That stance was contrarian a year ago. It's becoming the obvious move.

Smaller items: Challenger's final 2025 number landed at 54,836 AI-cited layoffs, third week running this figure has been the labor headline; full position in the job-security piece. The MIT Technology Review "great hype correction" piece argues 2025 was when AI "stopped being a fascinating abstraction and began meeting messy human reality." Mostly right and underplays one thing, the hype correction is concentrated at the consumer-product layer, but the infrastructure and capability layer kept compounding. Latent Space's WTF Happened in December 2025 recap frames the month as the most concentrated burst of capability in AI history. Five frontier-tier or near-frontier releases in a single month. GPT-5.2, Gemini 3 Flash, Claude 4.5 cycle, Mistral Large 3, GLM-4.7, and that framing is accurate. The European Commission proposed important changes to the EU's digital rulebook earlier in December; the early-2026 schedule is when this gets concrete. And a patient filed a class action on December 26 against Sharp Healthcare for using Abridge's ambient clinical-documentation app during his visit without consent. Early case of the consent-and-disclosure problem I've been flagging for ambient AI. The legal answer here will shape how every healthcare-AI deployment handles consent in 2026.

What 2025 told me, treating this as a partial year-end take with the formal version in the 2025 wrap-up. The infrastructure layer consolidated faster than the model layer. The story everyone told in January was "the model race." The story that actually mattered by December was the chip and infrastructure race. Nvidia-Groq is the bookend on a year that includes the OpenAI-AWS $38B deal and a dozen smaller hyperscaler-lab couplings. Models commoditize. Compute concentrates. The leverage moves accordingly. The open-model ceiling kept rising and the closed-model premium kept shrinking. January started with DeepSeek R1 making the case that the gap between open and closed was smaller than anyone thought. December ended with GLM-4.7 taking the top open-coding spot at a price point that makes hosted frontier hard to justify for most workloads. The trajectory is unambiguous. Run your own. And the labor displacement is real, accelerating, and being narrated dishonestly. 54,836 AI-cited cuts is the slice clean enough to count; the actual workforce restructuring around AI is bigger than that. Full read in the job-security piece.

What to watch next week: January 4, 2026, look-ahead edition. What I'm watching, what I think bends, which 2025 narratives I expect to break in the first quarter. The connecting thread across all three threads above: the principles I keep returning to, distributed over concentrated, sensitive data not in public AI, governance is the work, held up across 2025 as a frame for reading the news. They'll hold across 2026 too.