New Ask Claude or ChatGPT about any stock, using Bullpen. Bullpen in Claude or ChatGPT. 7 days free

Wednesday, November 18, 2026

Back to the front page

From X

Convequity on X

Steve Eisman's AI price-war call is early, because Chinese open-model releases are about to slow On 17 Nov Steve Eisman said Chinese open-weight models, which anyone can download and run, cost about $0.20 per million tokens and are taking share from Anthropic and OpenAI flagship models priced at $2 to $10. He expects a price war that neither lab has the balance sheet to fight. His call needs Chinese labs to keep releasing better open models. I expect those releases to slow sharply. US chip restrictions pushed Alibaba ($BABA), ByteDance and other Chinese firms to rent computing capacity overseas. About 75% of China's overseas AI compute is forecast to be in Southeast Asia by the end of 2026, per Institute for Progress estimates. My view is that most of Alibaba's training compute runs through that rented capacity. On 16 Nov the US Commerce Department released a rule under which data centers in Singapore, Malaysia, Thailand and Indonesia need a US license for advanced AI chips, including Nvidia's. Facilities already serving large Chinese customers have 90 days to apply, per CNBC. I think Alibaba is being cut off from that capacity. Hugging Face's release on 17 Nov shows how much the open-model community depends on Alibaba. It published an adapter, a small add-on trained on top of an existing model, that lifted the pass rate on 51 hard coding tasks from 29% to 63%, per its model card. The adapter works only on Alibaba's Qwen3.8-27B. A new base model needs a large training run, and the rule restricts the compute for those runs. DeepSeek chose the harder route of integrating its models with Huawei's Ascend chips. I expect a strong DeepSeek model after the middle of 2027, and open-model competition picks up again then. The timing makes me suspicious. Jensen Huang resigned as Nvidia CEO on 11 Nov, and the license rule came five days later. I read both as moves against open-source AI, and both help Anthropic around its IPO. I suspect whether Anthropic's stakeholders or their allies were involved. For the next two quarters, enterprises and AI startups will have fewer good new open models to choose from. Some will spend a little more on closed models, and I think Anthropic gains most.

Convequity on X

Anthropic has released Code Mode, a new way for Claude to carry out agent tasks. It works only through Claude Code on a Claude subscription or an Anthropic Console account. Anthropic's docs rule out Amazon Bedrock, Google Cloud and $MSFT's Foundry. Code Mode changes how Claude uses tools, and the model still runs on Anthropic's servers. Normally Claude makes each tool call as a separate model turn. With Code Mode, Claude writes one script that runs in the user's terminal, calls the tools and returns only the output Claude needs. Technically this is evolutionary. Anthropic has published no savings figure for Code Mode. Its closest number comes from the same technique on its API a year ago: 37% fewer tokens on complex research tasks. One early independent test of Code Mode, a single run, also measured 37% fewer input tokens, but with the same number of model requests and a slower finish. So Code Mode cuts tokens by about a third on tool-heavy tasks. The saving comes from Claude reading less tool output each step, not from fewer steps, so the dollar saving is likely smaller. That is useful and well short of a step change. The channel matters more than the technology. About 44% of Anthropic's revenue in the first nine months of 2026 came through Amazon, Google and Microsoft cloud accounts, per the prospectus. Those customers cannot use Code Mode. An enterprise that wants it has to buy Claude directly, and Anthropic keeps more margin on every workload that moves. Amazon and Google sit on both sides. As channels, they lose the middleman cut on each workload that goes direct. As holders of nearly 29% of Anthropic between them, they get part of it back through a more profitable Anthropic. I will watch two things: whether Code Mode reaches Bedrock or Google Cloud, and the cloud-account share of revenue against today's 44%.

Research Roundup Vol. LXI

Nvidia faced China workaround enforcement concerns as technology fell 6%; energy rose 4.5%, while the company raised eFY28 revenue growth…

NVIDIA Corp Earnings Update: Q3 2026

NVIDIA’s fiscal third-quarter 2027 results are its first earnings report since Jensen Huang’s resignation as CEO.

The Week Ahead

Micron’s November revenue guide will test whether an 86% gross margin reflects a structural memory shortage or a cyclical peak, while…

CBOE · long

Cboe Global Markets INC

Cboe Global Markets trades near $232, with a roughly $24.3 billion market cap, net cash and a 16-17x forward P/E.

Hindsight Capital

Hindsight Capital tests crisis-era investment decisions using anonymized company financials, including TransDigm, Halma and Texas Pacific…

Publications we follow

Every publication on this site, hand-picked.

130 publications

Newsletters 110

X accounts we follow

Every account on this site, hand-picked.

91 accounts

Investor letters we follow

383 firms

Sectors

Ordered by how often our editor reaches for them.