LTX-2, NVIDIA Rubin and ChatGPT Health: January 2026 AI News

In this January 2026 AI news roundup, Matthew Berman starts with LTX-2, an audio-video generation model whose weights and development tools were released for local use. The release includes full and distilled weights, LoRA adapters and a training framework with multimodal pipelines and multi-GPU support. Availability and performance depend on the particular files, license, hardware and configuration.

Berman says the model is optimized for NVIDIA hardware, including the RTX 5090. He highlights controls for starting and ending frames, keyframes and camera movement, and Lightricks' claim of clips up to 20 seconds with synchronized audio, 4K output and up to 50 frames per second. These are provider-stated capabilities, not proof that every setting runs at those maxima on every consumer GPU.

He points to an Artificial Analysis open-weight leaderboard shown in a January 7 post and selected video demos. Rankings are snapshots of a particular evaluation, and polished examples do not establish performance on every prompt. The model weights and related code were made available through the developer's repositories, subject to their respective terms.

At CES, NVIDIA announced its Rubin platform for data centers rather than consumer PCs. NVIDIA claimed up to a tenfold reduction in inference token cost and fourfold fewer GPUs for training mixture-of-experts models compared with Blackwell. Those are NVIDIA's upper-bound comparisons under its assumptions, not independent guarantees for all workloads.

In a keynote excerpt, NVIDIA CEO Jensen Huang says the system can use 45°C inlet water without data-center water chillers. Berman describes a 16-rack configuration containing 1,152 GPUs and says NVIDIA reported Rubin in production, with partner products expected in the second half of 2026. Chiller-free cooling still uses a cooling system; it does not mean the hardware requires no water or heat rejection.

Berman next discusses OpenAI's announcement of ChatGPT Health, a dedicated space for health and wellness conversations with optional connections to health records and apps. He says he already uses ChatGPT to help understand test results before speaking with his doctor and is interested in adding personal data. OpenAI described the feature as supporting medical care, not diagnosing or treating patients.

He recognizes that people may be uncomfortable sharing health information with a third party. He imagines an alternative in which Apple could use health data already held on a device for local AI assistance, keeping the information on the device. This is his proposed possibility, not an announced Apple feature in the video.

At launch, OpenAI described a waitlist and initial eligibility for Free, Go, Plus and Pro users outside the European Economic Area, Switzerland and the UK; some health-record and app integrations were US-only. OpenAI said Health conversations were separate and not used to train its foundation models. Berman had not yet tried the feature, and availability and privacy terms were presented as a January 2026 snapshot.

Berman discusses a social-media claim that memory shortages could push an RTX 5090 from $2,000 to $5,000. He also reads the accompanying note that neither NVIDIA nor AMD had confirmed those specific increases. The amount is speculation from the time, not an announced price or a reliable current shopping recommendation.

A reporter at CES asked Huang whether older GPU production or lower-memory configurations might ease supply pressures. Berman recounts Huang saying it was possible and worth investigating, including the engineering work required to bring newer AI features to previous GPUs. That response was not a commitment to restart production or a confirmed price change.

Berman moves to NVIDIA's Alpamayo family of models, simulation tools and datasets for autonomous-vehicle research. He describes vision-language-action reasoning and a Mercedes demonstration, and sees the approach as potentially useful for carmakers. NVIDIA's claims about reasoning and long-tail driving situations are research and product claims, not proof of safe deployment in every condition.

He says synthetic driving scenes can support training and simulation and contrasts the approach with systems built using extensive real-world driving data. His suggestion that synthetic data eliminated the need for real driving data is too broad: NVIDIA also released real-world driving data. Its January announcement describes an open dataset with more than 1,700 hours, not evidence that the model was trained exclusively on that amount.

NVIDIA described Alpamayo 1, AlpaSim and its Physical AI Open Datasets as resources for developing systems that perceive and act in challenging scenarios. Berman says the models and tools can be downloaded and adapted, while predicting that incumbent manufacturers may struggle to implement them. The announcement and a selected demonstration do not establish that a complete, validated autonomous vehicle can simply be installed by any car company.

Berman describes the demonstration as using camera input to produce driving actions and compares the sensor approaches of Tesla and Waymo. He mentions his own Tesla ride without disengagements, which is an anecdote rather than a safety comparison. Sensor configurations, operational domains and safety performance cannot be inferred from this brief demo alone.

Finally, Berman cites a January 7 Wall Street Journal report that Anthropic was seeking to raise $10 billion at a $350 billion valuation. This was a reported fundraising plan at the time, not a confirmed completed round. His claim that Anthropic was nearly profitable is his assessment and is not established by the quoted fundraising report.