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

Friday, November 20, 2026

Back to the front page

Grand Expectations

Grand Theft Auto VI launched to a 99 Metacritic score after estimated $1.5 billion to $2 billion development and marketing spending, raising questions about Take-Two’s returns.

From X

Convequity on X

Gemini 4 Neon ranks first on coding benchmarks. After testing the Gemini 4 models, I would not replace Opus with them for engineering work Google ($GOOGL) launched Gemini 4 Neon on 18 Nov at $3 per million input tokens and $15 per million output tokens. It ranks first on the most-cited independent coding-agent index with a score of 58.6, and X is arguing whether it dethrones Claude. Neon is a real step up for Google. It is far better than Gemini 3.8 Flash, and for the previous few quarters Google users only had a Flash-tier model to work with. Its writing is very impressive, and so is its world knowledge. In my testing currently, Gemini 4 behaves more like GPT-6.2 Astra than like Opus in engineering work. I see the same two problems I see with Astra. It over-engineers, which means it builds more than the task asked for. And it follows the instructions or the agent guide to the letter without using its own judgment, so it can be single-minded in a way that does not look intelligent. Benchmarks do not capture this. Google's own launch blog shows how much a coding score depends on the setup. On the SWE-sweep test, the same Neon model scores 12.8% in a basic harness and 39.0% inside Antigravity, Google's own agent product. A benchmark score records whether the task was completed in that setup. It does not record whether the model built only what was needed, or whether it knew when to depart from its instructions. I call that taste, and we need a benchmark or some other way to verify it. I have not seen a dramatic take-up in Gemini usage or an oversubscribed Antigravity, and I think this is the reason. OpenAI and Google have the compute and the know-how to train a larger model and reach the top tier of the benchmarks. In daily use their models are still not comparable to Anthropic's. Anthropic is now public as $ANT, and I expect its competitive position to stay strong for the next two quarters.

Convequity on X

The new US rule on Chinese access to overseas GPUs could push GPU rental prices down. That hurts GPU cloud companies and helps Anthropic The US Commerce Department rule that took effect on 19 Nov requires a US licence to transfer advanced GPUs to any data centre in the world that provides more than a set amount of compute to companies headquartered in China. The threshold is 12 million TPP, the US government's measure of chip performance, which is reportedly about 300 of the most advanced GPUs. Existing providers have 90 days to apply for a licence. The first casualty is already visible. Datasection, a Japanese neo-cloud, meaning a company that rents out GPU capacity, cut its fiscal-year revenue guidance by 89% after a large order was cancelled. Reports say its capacity was in practice a Tencent data centre running Nvidia ($NVDA) Blackwell chips, and that it still has unpaid GPU purchase commitments. Most of the commentary asks whether the loophole is closed. I think the bigger effect is on GPU rental prices. The capacity that was reserved for Chinese customers still exists, and its operators still have GPU bills to pay. They have to find new buyers in a very short period, and the place to find them is the public spot market, where GPU capacity is rented at the current price instead of on a long contract. My estimate is that this capacity is 10% to 15% of global GPU cloud supply. And it can be a vast majority of the spot rental you can get for a large compute capacity which is a rare opportunity. That much capacity looking for buyers at once pushes spot prices down. In the short term this hurts Western neo-clouds such as CoreWeave ($CRWV), IREN ($IREN) and Nebius ($NBIS), and even SpaceX. Lower spot prices reduce what they can charge for capacity that is not yet under contract. It is good news for the buyers. Anthropic ($ANT) and other AI labs have been suffering from rising compute costs. If Anthropic locks in this capacity, its compute cost falls over the coming quarters and its gross margin improves. I also expect open-source models to lose ground over the next two quarters. Cheaper compute and weaker open-source competition arriving together make this a very good period for Anthropic. Watch for more guidance cuts from non-US neo-clouds before the 90-day licence window closes in mid-February 2027.

Problems at AMD (AMD)

AMD has committed up to $161.1 billion in customer equity rights and $33.1 billion in guarantees, leases and investments, while holding…

Thursday Tailwinds #15: Nov 19, 2026

India’s 236 GWh battery target boosts local electrolyte-salt makers, while Mirion’s radiation monitors and Overwatch’s passive drone…

Publications we follow

Every publication on this site, hand-picked.

130 publications

Newsletters 110

Seeking Alpha 98 of 301

X accounts we follow

Every account on this site, hand-picked.

94 accounts

Investor letters we follow

383 firms

Sectors

Ordered by how often our editor reaches for them.