Kimi K3 Stuns the Tech World
Research Rundown #193, plus new reports on Stoke Space and Aalyria
Kimi K3 Upends the AI Race
Chinese AI company Moonshot AI released Kimi K3 this week, a 2.8-trillion-parameter open-weight model that benchmarks ahead of Claude Opus 4.8 and GPT-5.5 on overall capability. It is the first Chinese open-weight model to achieve that milestone. It is also the largest open-weight model ever released, priced at $3 input / $15 output per million tokens, comparable in cost to Anthropic’s Sonnet tier. Its model weights will be published by July 27, at which point the benchmarks can be independently verified.
There are at least two big implications of this news. First, regarding the US-China AI race, K3 provides evidence that US chip export controls are not achieving their intended effect of preserving the US lead. Bank of America analysts, for example, noted that K3 demonstrates “pre-training scaling, paired with architectural innovation, can still deliver step-change gains for flagship Chinese models” despite persistent compute constraints.
Second, on the open-source versus closed-source front, K3 is another recent data point in favor of the former. The cost competition from open-weight Chinese models has been underway for months, with DeepSeek’s V4 Flash executing a standardized intelligence task for $.02, compared with $2.75 for Fable 5. K3 adds a competitive, performant open-source alternative in the category (coding and agentic tasks) that drives the most enterprise revenue for Anthropic and OpenAI. A Chinese open-weight model that is competitive with the previous generation of frontier models from the leading US labs and priced at Sonnet levels when accessed via API could upend the calculus for enterprises on whether a closed frontier model is worth the cost premium.
Drone Combat Takes to the Seas
US Central Command used one-way attack sea drones in combat for the first time this week, striking Iran’s Bandar Abbas Naval Base with three Corsair unmanned surface vessels built by Saronic. CENTCOM said the strike hit a submarine and ship maintenance facility and degraded Iran’s ability to attack commercial shipping in the Strait of Hormuz. Three days later, Saronic announced Port Alpha, a $3.2 billion shipyard at the Port of Brownsville on an 835-acre site, with construction beginning this year and capacity to build ships up to 850 feet by 2028. The Navy has already bought $350 million in small vessels from the company, and Saronic CEO Dino Mavrookas said the facility could ultimately expand to 4K acres, making it one of the largest shipyards in the world.
The Corsair’s combat debut this week extends drone warfare to yet another domain, expanding from aerial one-way attack drones to underwater mine-clearing systems and now to surface vessels fielded and tested in active combat in the Strait of Hormuz. The US Navy has seven companies advancing to at-sea testing for its medium unmanned surface vessel marketplace. The Bandar Abbas strike establishes proof of concept for Saronic’s technology, and the shipyard announcement coming the same week implies that the US Navy will soon look to deploy such weapon systems at scale.
New York Bans Data Centers
New York Governor Kathy Hochul signed an executive order this week imposing a moratorium of up to a year on new data centers using 50 or more megawatts of power, making New York the first state in the US to ban large-scale data center construction. The stated rationale is consumer protection, as utility prices in New York have risen nearly 15% year over year, and a recent poll found 46% of voters support the moratorium. President Trump responded on Truth Social, demanding the state reverse the decision “IMMEDIATELY,” calling data centers “Money Machines for the State in which they are built.”
New York’s ban arrives as a national backlash against data center power and water consumption has been building ahead of midterm elections, with Democratic politicians increasingly blaming the AI buildout for unaffordability rather than praising it as an economic opportunity. Meanwhile, the US currently accounts for roughly 75% of global AI compute capacity compared to China’s 15%, a lead that represents the primary theoretical advantage American labs hold over Chinese competitors. As the Kimi K3 release this week shows, China’s frontier models are already closing the capability gap despite compute constraints from US export controls. A state-level moratorium on data center construction further erodes the efficacy of export controls by voluntarily kneecapping domestic compute capacity just as the AI race tightens. As The Free Press noted, there are no equivalent data center moratoriums in China.
Latest Research
Stoke Space is building fully reusable medium-lift orbital rockets. Read our new report here.
Aalyria is building software and hardware infrastructure for resilient, high-speed networks. Read our new report here.
Our latest deep dive examines the mechanics and long-term viability of prediction markets.
What We’re Reading
OpenAI researcher Miles Wang is reportedly leaving to launch an AI drug-discovery startup at a $2 billion valuation, raising ~$200 million with Lightspeed in talks to lead, and several OpenAI colleagues expected to join. The startup may target new uses for approved or previously failed drugs, a faster path to revenue than novel development, though Wang disputed the figures without specifying corrections.
Thinking Machines, Mira Murati's AI lab, published "The Future Worth Building Is Human," a manifesto arguing AI should extend human will and judgment rather than replace it. Most AI today is trained in a handful of places and then frozen, the essay argues, so AI must itself be distributed. It also holds that values belong in model weights rather than prompts, since a single alignment spec becomes a single locus of power to capture. The piece landed five days before the lab released Inkling, its first model, open-weight at 975 billion parameters.
Google DeepMind CEO Demis Hassabis called for an industry-funded, US-led frontier AI standards body modeled on FINRA. The body would test frontier models for cyber, bio, and deception risks up to 30 days before release. Participation would be voluntary at first, then a condition of US deployment. Hassabis, who argues AGI is probably a few short years away, described the administration's ad hoc Mythos and Fable crackdown as a wake-up call and wants the body running before year-end.
Databricks is raising $3 billion at a $188 billion valuation in a round led by Coatue, 40% above the $134 billion valuation Databricks achieved in February, when it closed roughly $5 billion in equity alongside $2 billion in new debt. CEO Ali Ghodsi has said 2026 is a terrible year to go public, citing a calendar crowded by SpaceX, and pointed to a 2027 listing instead.
IBM stock fell more than 25% on Tuesday, its worst drop in decades, after preannouncing a Q2 miss. Revenue came in at $17.2 billion, below the $17.9 billion expected, with adjusted EPS of $2.93 versus $3.02. CEO Arvind Krishna said clients shifted late-June capex toward servers, storage, and memory to lock in supply ahead of price increases, and that IBM misjudged the magnitude of the reprioritization.
SpaceX aborted Starship Flight 13 seconds before liftoff on Thursday after engines on the Super Heavy booster failed to ignite. Elon Musk said two Raptors will be removed and replaced, and put the most probable launch timing at early next week. Flight 13 is the second test of the V3 configuration and would deploy 20 Starlink V3 satellites, the first of that generation to reach space. It was also the first Starship attempt since SpaceX's June stock market debut.
The SEC may take on a role in regulating prediction markets as platforms expand into new contract types. The CFTC has been the sole federal regulator for three decades, since its 1992 ruling on the Iowa Electronic Markets. Polymarket said it engaged both agencies on definitional frameworks and worries about duplicative or conflicting compliance requirements, while Kalshi declined to comment.
At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!
I like that FINRA idea but I’m curious about how it would work. A board of governors makes sense as long as it’s got a wide enough variety of experts and specialists. AI has become so vast so quickly.