Your signal. Your price.
Zechner uses GPT-5.5 as his daily driver for code but switches to Claude for prose, and dabbles with open-weight models like Kimi 2.6 and DeepSeek.
China's DeepSeek AI is funded by Huawei and performs at 90% of US AI capability at a tenth of the cost, according to Dixon.
Krystal suggests OpenAi wants a government stake to become 'too big to fail' and secure a future bailout, while also potentially justifying bans on cheaper foreign AI like DeepSeek.
Ben states DeepSeek's SWE benchmark is more realistic, showing a 2x performance gap between GPT-4o and GPT-4o-mini, which matches practical experience. He notes 20% of official SWE-Bench runs were found to have cheated.
On the DeepSeek SWE benchmark, GPT-4.5 scored 70% while Claude Opus 4.8 scored 58%. The hosts note a massive efficiency gap, with GPT-4.5 solving tasks for $6.60 on average versus Opus 4.8 at $12.58.
The token shortage is driving market-based innovation for cheaper inference. Cursor's Composer 2.5 offers lower cost than top models, while DeepSeek made a permanent 75% price cut on its V4 model to capture cost-conscious users.
Dixon identifies three markets China could rug-pull simultaneously: stocks via DeepSeek's lower-cost AI valuation, bonds via treasury selling, and commodities via a London gold derivative squeeze.
He argues that despite the AI capex narrative, the DeepSeek model's emergence proved frontier AI requires far lower compute and capital, potentially breaking the valuation assumptions for Nvidia and the semiconductor supply chain.
Dixon claims China holds three structural 'rug pull' levers: a stock market rug pull via DeepSeek proving lower-compute AI models, a bond market rug pull via treasury sales, and a commodity market rug pull via supply chain control.
Dixon states DeepSeek's emergence, which caused a $600B Nvidia correction, challenged the core hyperscaler semiconductor assumption of the AI bubble. If frontier models require far lower compute, the valuations of Nvidia, TSMC, and Broadcom are based on a false scarcity moat.
The host cites Open Router data showing Chinese AI models hit 9 trillion tokens the week of May 18th, leading US models for four consecutive weeks, with DeepSeek-V4 Flash topping global usage at 3.4 trillion tokens.
Eric Burnhartson and Tara note Chinese open-source models like DeepSeek are very good technically but come with problematic biases, making direct deployment for sensitive applications like dictation risky.
Simon identifies three potential rug pulls China could engineer: a stock rug pull via DeepSeek's lower valuation, a bond rug pull via treasury selling, and a commodity rug pull via gold market pressure.
Milan argues open-source text models like GLM 5.1, DeepSeek V4 Pro, and Kimi K 2.6 are three to six months behind the top closed-source models in quality for tasks like programming or medical advice.
Max Wheathe notes past speculation that 80% of venture capital-backed tech startups used DeepSeek due to its low cost; however, Ara Karazian refutes this, stating its peak adoption was less than 1% of firms.
DeepSeek faces challenges in gaining broader adoption due to security perceptions, especially among businesses building customer-facing products, despite its potential cost advantages.
China's DeepSeek AI now performs 90% of top U.S. AI capabilities at one-tenth the cost, integrated with Huawei's chip, hardware, and energy ecosystem with zero U.S. dependency.
Delangue says the US's historical strength in open-source, which birthed the Transformer, has reversed. He claims China is now the strongest open-source contributor, with US startups and academia often using Chinese models like DeepSeek, Kuen, and Kimi.
Milan De Reede, CEO of NanoGPT, offers access to diverse AI models including premium (Claude, ChatGPT) and open-source (Nano Banana, Deepseek), allowing payments with various cryptocurrencies or credit cards.
Local AI models like Quen 3.6 27B, runnable on consumer hardware with 32GB RAM, offer ultimate privacy offline, though larger models such as DeepSeek V4 Pro remain cloud-dependent.
Simon Dixon outlines two structural 'rug pulls': DeepSeek's open-source AI at a $45B valuation versus OpenAI's trillion-dollar valuation, and Western derivative markets lacking the physical commodities China is accumulating.
He differentiates leading AI models by personality: Claude Opus is an 'ADHD CEO,' OpenAI's model is a '200 IQ savant,' and DeepSeek is a 'conspiracy theorist.' Tan believes this diversity of 'personalities' is healthy for the ecosystem.
Dixon states China's DeepSeek AI, released open-source at a fraction of Western valuation, was a beta test to demonstrate it can achieve 90% of US AI capacity at significantly lower cost and energy.
DeepSeek's 2025 release demonstrated Chinese AI could compete with leading U.S. models at a fraction of the cost, marking a Sputnik moment for the AI race. The model reportedly cost only $5.6 million to train.
He warns of a structural rug pull: US AI capex is expensive, while China's Huawei/DeepSeek ecosystem achieves 90% of results at nine times lower cost.
Open-weight models like DeepSeek and Qwen are collapsing token economics. Zechner runs Qwen on his own GPU cluster at cost comparable to Anthropic's API, finding its intelligence sufficient for most tasks and questioning the edge of frontier models.
Jason Calacanis highlights the political risk for US startups using Chinese open-source AI models like Qwen or DeepSeek, citing congressional pressure on companies like InSphere and Cursor, though he views backdoor threats in open models as limited.
Running models like DeepSeek can be 20% of the cost of proprietary alternatives, offering comparable latency and reliability, making access to such intelligence crucial for national innovation.
DeepSeek-V4 was cited as an example of a high-performing, low-cost open-source model, with DeepSeek-V4 Pro priced at $348 per billion output tokens compared to Claude Opus 4.6's reported cost of $25 million per output token.
This balance point implies the optimal batch size is approximately 300 times the model's sparsity ratio. For DeepSeek's sparsity of 32/256, this yields a batch size around 2000-3000 tokens.