Kimi K3 & AI’s Price War, What Happened To Google?, OpenAI’s Partner Trouble

Summary

Alex Kantrowitz and Ranjan Roy focus on Moonshot’s Kimi K3 as a DeepSeek-style wake-up call: a large open-weight model that performs near frontier levels while undercutting some top U.S. models on price. They argue the real shift is from raw model intelligence to model interoperability, harnesses, and product experience, because if several firms can reach similar capability, APIs and inference margins start to commoditize. The discussion then turns to Google, where Bloomberg-reported delays to Gemini 3.5 Pro are framed as a sign that Google is struggling to adapt to the coding and autonomous-agent phase of the AI race, despite strengths in multimodality and distribution. The episode closes on Apple’s lawsuit over alleged OpenAI trade secret theft and a broader concern that OpenAI is accumulating enemies rather than allies, which could become a strategic weakness in a business that still depends on partnerships.

Notes

• The episode frames Kimi K3 as a major catalyst for AI price compression and frontier-model commoditization.
• Kimi K3 is described as a 2.8 trillion parameter challenger from Moonshot.
• Bloomberg’s framing is that Kimi K3 rattled markets and hit AI and semiconductor names.
• A downstream worry is that U.S. firms shifting to Chinese models could reduce Anthropic investment and chip demand.
• Alex says Kimi K3 lands near GPT 5.6, Opus 4.8, and other top models on major benchmarks.
• Ranjan calls Kimi K3 a DeepSeek-like moment because it challenges the moat value of frontier models.
• He says buyers increasingly want the best model for each task, not one frontier label.
• Alex pushes back that Kimi is not dramatically cheaper and still does not beat every frontier model.
• Ranjan highlights the pricing as the core strategic signal: $3 per million input tokens and $15 output tokens.
• He says Kimi is 40% cheaper than GPT 5.6 and 70% cheaper than Fable.
• Ranjan distinguishes Moonshot’s approach from DeepSeek by emphasizing quality plus affordability.
• Alex notes Kimi K3 will be open weight, so sophisticated buyers could download and self-host it.
• He says that combination of open-weight Chinese models and cheaper rival models could commoditize the API layer.
• Alex quotes a thesis that Kimi K3 could hurt OpenAI and Anthropic but help the rest of the AI stack.
• That thesis extends to frontier labs acting as monopsonies over power, data centers, semiconductors, and hyperscalers.
• Alex relays the claim that open-source models still require similar compute to run as closed models of similar size.
• The broader implication is that lower model-layer margins can shift value to infrastructure and software.
• Alex argues that if several firms reach similar intelligence, the model layer becomes a price war.
• Ranjan says product and UI still matter, because the interface can make AI feel more intelligent.
• Alex says if intelligence is abundant, winning means being the best AI product builder.
• Ranjan says Anthropic’s AGI/superintelligence story weakens if the company is mainly a good product shop.
• Alex counters that Anthropic has an advantage when model and product are tightly integrated.
• Ranjan calls Claude Code an incredible product even though it is only a command-line tool.
• Ranjan warns that customers may avoid Anthropic if they think it will reuse their data to build competing products.
• Alex says Kimi K3 makes switching away from Anthropic easier because the model is similarly capable and customizable.
• Alex invokes Satya Nadella’s “reverse information paradox” as a warning about AI vendor lock-in.
• He says the seller can learn more about the buyer over time as usage deepens.
• Ranjan says the conversation has swung back toward data training and proprietary data flywheels.
• Alex says social-media skeptics pointed to Kimi running on Huawei 14nm hardware against Blackwell-heavy systems.
• He also argues that infrastructure owners could still win even if model labs failed.
• Alex extends that to a circular-financing logic where infrastructure providers get paid either way.
• David Sacks is quoted saying Kimi K3 is number one on front-end code arena and near frontier elsewhere, while U.S. regulation is slowing America.
• Alex criticizes the irony of Sacks complaining about interventionism after the Fable ban.
• Alex says controlled rollout can make sense for risky cybersecurity-capable models, but politicized bans do not.
• Bloomberg reports Google’s Gemini 3.5 Pro is months behind schedule because coding performance needs improvement.
• The reported delay reflects Google’s need to coordinate many stakeholders across Search, Maps, YouTube, and more.
• Ranjan says Gemini felt caught up for a while but has stalled over the last seven or eight months.
• Alex says the market shifted from cheap parity models to autonomous coding systems.
• He suggests Google may have over-invested in Flash-style smaller models instead of unifying around bigger frontier systems.
• Ranjan says Google’s coding work is fragmented across Cloud, DeepMind, and Android.
• Alex says internal purists who insist important code be human-written are a strategic drag.
• Ranjan recalls Sundar Pichai saying a large share of Google code is written by AI, and Alex corrects the number to 75%.
• Alex says Google should never be short on compute if AI is a priority.
• Ranjan responds that Google Cloud’s size and autonomy make resource reallocation politically hard.
• Ranjan says Google still seems strongest in image and video multimodality.
• Alex describes the Apple v. OpenAI trade secret dispute as an unusually clumsy espionage case.
• Ranjan agrees the alleged espionage effort is extraordinarily dumb.
• Alex says Apple has sent preservation letters to dozens of OpenAI employees and may take the case into discovery.
• Alex says OpenAI’s strategic problem is that it keeps turning partners into enemies.
• He concludes that tech companies need allies, and OpenAI’s loss of friends could become costly.