HiDream-O1-Video debuts fourth on Artificial Analysis
A Chinese startup put a video model on the global boards at the top on release day. Separately, the US chip buildout is running straight into the thing money cannot buy fast enough: technicians.
HiDream.ai released HiDream-O1-Video-1.0 (HD-V1) on September 15, and it landed fourth worldwide on Artificial Analysis's Image-to-Video leaderboard for models with audio — an Elo of 1176, behind MiniMax H3 Max (1195), MiniMax H3 (1181) and Google's Gemini Omni Flash (1177), and ahead of ByteDance's Dreamina Seedance 2.0. The same model placed eighth on Arena.ai's image-to-video board. HD-V1 is native omni-modal in the strict sense: text, image and video go in, 5–20 second 1080p comes out, and audio is generated jointly with the picture instead of being spliced on afterwards. HiDream calls the training approach Diffusion Reinforcement Learning with a multimodal reward model; the practical upshot on the demo reel is a music stage where the lip sync survives a deliberately mangled lyric.
The more interesting claim is physical. A hands-on test by the Chinese outlet 智东西 ran the model against scenes where one frame has to imply the next: a screw driven into a plank with both hands on the tool, an NBA buzzer-beater three generated from a single still reference, and an eight-lane Olympic 100m final where the field has to separate and then converge at the line. The model held the runners' relative positions and their spacing through the race — the kind of continuity that separates a video generator from a world model, which is exactly the framing HiDream is selling: one native multimodal base with image, video, interactive and embodied models co-evolving from it. We covered its first swing at that thesis when HiDream-O1-Embodied topped RoboColiseum's robustness chart. Alongside the launch, HiDream closed a Series C+ round from Xinwei Capital, Jiaozi Capital and ICBC Capital; the size was not disclosed.
The US chip buildout will be short as many as 157,000 skilled workers by 2030, according to a McKinsey and SEMI Foundation analysis — and the companies building the fabs are saying it out loud. "I'm concerned," Samsung's semiconductor executive vice president Jon Taylor told CNBC from Austin. "We just don't see that there's enough technical people in the pipeline." The arithmetic behind that: only 3% of US engineering graduates who take engineering roles go into semiconductors, and 73% of chip employers report significant difficulty filling engineering roles. Samsung's two fabs in Taylor, Texas — part of a $35 billion state buildout expected to create roughly 3,500 jobs — start production later this year, and it is flying in South Korean staff to commission equipment nobody in Texas has run before.
The pay gap explains the flow in the other direction. Asian fabs have handed out bonuses above $500,000 during strike negotiations, while US semiconductor salaries run roughly $127,000 to $187,000. The patch is domestic pipeline: Purdue now enrolls about 2,500 students a semester in chip courses, Arizona State converted an old Motorola fab into a cleanroom with $200 million from Applied Materials, and TSMC is filling 6,000 roles at its first three Arizona fabs partly from interns. Our read: the memory shortage that has Nvidia and AMD scrambling is a staffing shortage wearing a supply-chain costume. Capex buys fabs; it does not buy the people who can run them.
What to watch: whether the federal workforce fund scales with the fab buildout, and whether Samsung's Taylor fabs hit their start date with borrowed Korean crews.
If the constraint is technicians rather than capital, should CHIPS-era money shift toward training pipelines instead of more fabs? Tell us in the comments.
Sources: 智东西 Zhidx · Sina Finance · Artificial Analysis · AI Market Watch · CNBC · McKinsey · SEMI Foundation / NNME
