In partnership with
In Today’s Issue:
🎨 Qwen-Image-2.1’s open-weight challenge
💵 Anthropic’s IPO and model-release decisions
📈 Astra’s early spending lead on Vercel
💰 OpenAI’s trillion-dollar funding talks
✨ And more AI goodness…
⚡ The Signal
Anthropic is weighing a new model release while preparing to go public and calling for a slowdown in AI development.
Its reported revenue growth gives investors a reason to keep backing it, while Astra’s early spending lead on Vercel helps explain the competitive pressure. OpenAI’s cash forecasts add another concern: even rapid growth can leave a lab needing more funding. I’m watching how patient investors will be if releases slow down while computing bills keep rising. Qwen, meanwhile, has released image-model weights that developers can download and test themselves.
All the best,
Kim Isenberg
Reported annualized revenue and 2026 projection. Source: The New York Times, September 18. Graphic: Superintelligence.
⚖️ Anthropic’s IPO Push Meets Its Safety Warnings
Anthropic is pressing ahead with its IPO despite Dario Amodei’s call for stronger AI guardrails, The New York Times reports. Sources expect its annualized revenue run rate to exceed $100 billion this year, up from $65 billion in July, which investors see as support for a $2 trillion valuation. Annualized revenue shows what a sales pace would amount to over a full year, rather than revenue already earned.
👉 tl;dr: Anthropic’s growth is keeping its IPO moving while investors weigh the risks its own CEO is warning about.
💵 Anthropic Targets November for IPO
Anthropic is now targeting November for its stock-market debut, a month later than investors had expected, according to The Wall Street Journal. Some advisers say the extra time would let it present third-quarter results showing how the business has held up since OpenAI launched Astra. People familiar with the timing say the decision came before the latest public safety debate, and the date could still change. 👉 tl;dr: Anthropic’s IPO is penciled in for November, potentially giving investors a clearer picture of its business after Astra’s arrival.
🧠 Anthropic Could Release a New Model
Anthropic is considering a new model release before its planned IPO to counter OpenAI’s Astra, Reuters reports, citing three people familiar with the matter. The Claude maker is evaluating the model’s safety and declined to comment; the report gives no release date or model name. 👉 tl;dr: Anthropic may answer Astra with a new model before asking public investors to buy its shares.
🎬 Watch This
Jensen Huang tells CBS: “We should go as fast as we can, but not faster than we should.” He also rejects unsafe releases. I agree with the principle: move quickly, with safety setting the limit on what ships. His confidence makes this interview a useful counterpoint to Amodei’s warnings.
VB’s post points to a Linux preview of the ChatGPT app, with projects, local files and Codex shown in the documentation screenshot. A welcome development for people who do their work on Linux.
A more specific Anthropic rumor: @lyraxana claims Opus 5.5, tested as claude-wafer-eap, is planned for Tuesday, September 22. This remains an unconfirmed claim from the post; the Reuters reporting above does not establish that model identity or date.
Qwen Releases Its 7B Nano Banana Challenger
The Takeaway
👉 Downloadable weights: generation and editing in one model.
👉 7B visual generator, with a separate text encoder.
👉 60.28 on Qwen’s test, ahead of every model with a disclosed size.
👉 Up to ten references, plus transparent-image output.
Qwen released Qwen-Image-2.1 on September 20, with downloadable weights for a 7B visual generator that creates and edits images. On Qwen Image Bench, it scores 60.28 points, ahead of Nano Banana 2.0’s 59.82 and every model in the chart whose size is public. Two larger models trail it: FLUX 2 Max (32B parameters, 55.33 points) and Hunyuan Image 3.0 (80B parameters, 50.81 points). Six closed models score higher, led by GPT Image 2.5 Sunburst at 67.01. This is Qwen’s own evaluation. Developers can now test the released weights against their own images.
Qwen Image Bench overall scores and disclosed parameter counts. Source: Qwen.
Feed it a person, an outfit and a few accessories, and Qwen can combine them into a new picture. It accepts up to ten reference images, supports targeted edits and can generate transparent backgrounds. Qwen’s example below puts five separate references into one outfit image. For a design workflow, keeping the person and products consistent across edits is a really useful test.
Five references assembled into one outfit image. Model-generated example supplied by Qwen.
The Qwen Research License Agreement permits noncommercial research and evaluation. Commercial use requires a separate license, including when turning a successful experiment into a business product. The 7B figure covers the visual generator, which works alongside a separate Qwen3-VL-8B text encoder. As it refines a picture, it reuses information from the prompt and reference images instead of processing that information again at every step. Diffusers and ComfyUI supported it from day one, giving developers established tools to start testing those image workflows.
Why it matters: Developers can evaluate an image model against their own references and inspect how it works. For a design workflow, consistent products and controllable edits are what would make it useful. It also outperforms Googles Nano Banana 2.0 on key benchmarks.
The 10 Best AI Stocks to Own in 2026
AI is moving from experiment… to essential.
Every major industry is integrating it.
Every major company is investing in it.
By late 2025, AI was already an $800B market — growing at a pace that could push it well beyond $1 trillion in the years ahead.
Cloud infrastructure is scaling fast.
AI-enabled devices are multiplying.
Automation is becoming standard.
But here’s the real question…
When trillions flow into this transformation — which stocks stand to benefit most?
Our new report reveals 10 AI stocks positioned across the backbone of this shift — from the companies powering the infrastructure… to those embedding intelligence into everyday systems.
If you want exposure to one of the defining growth trends of this decade, start here.
Cumulative AI Gateway spend over each model’s first 12 days. Source: Vercel, September 17 update.
The chart: On Vercel’s AI Gateway, spending on GPT-6 Astra passed Fable 5.1 on day four and reached twice Fable’s total by day 12. Each line starts at its own launch date, with Fable’s twelve-day total set to 100.
The lesson: Astra attracted more early spending on this service. Vercel says both models launched at the same price. That lead helps explain the competitive pressure behind Anthropic’s possible model release in the Reuters story above.
The caveat: Vercel estimates spending from published list prices, so actual bills may differ. This is one gateway, influenced by its customers and workloads. It tells us how these launches started there, rather than establishing global market share or which model is smarter. The launches also happened on different calendar dates.
💰 OpenAI Projects $278B in Cash Burn Amid Trillion-Dollar Talks
⚡ Bottom line
OpenAI is discussing trillion-dollar valuations while projecting $278 billion in cash burn across 2026 to 2030.
💡 Why it matters
Even its projected surge in revenue leaves OpenAI needing large amounts of outside capital to keep growing.
🔎 What it means
A successful model launch can strengthen the fundraising pitch while adding to the bill for serving customers.
OpenAI’s next funding round could put a trillion-dollar price on a business that still expects years of heavy cash burn. Reuters reported September 18, citing a company presentation reviewed by the Financial Times, that OpenAI projects $278 billion in cumulative negative free cash flow from 2026 through 2030. That measures cash consumed across five years, rather than the company’s value or one year’s loss.
Projected cumulative negative free cash flow, 2026 to 2030. Source: Reuters / FT, September 18. Graphic: Superintelligence.
Investors and OpenAI are starting from different prices. Reuters’ September 15 report described early, investor-initiated talks at about $1.2 trillion. The New York Times reported September 16 that investors proposed that valuation, while OpenAI wanted at least $1.5 trillion. No agreement had been reached. Its completed March 31 financing brought $122 billion in committed capital at an $852 billion post-money valuation, the company’s value including the new investment.
Completed March post-money valuation versus the reported investor proposal and OpenAI’s minimum target. Sources: OpenAI, March 31; NYT, September 16. Graphic: Superintelligence.
The plan assumes annual revenue grows from $36 billion in 2026 to $350 billion in 2030, according to the Reuters/FT report. Those are forecasts. Even with that growth, the company expects to consume cash across the five-year period as it pays for computing power and infrastructure. More sales can increase both revenue and the cost of serving customers.
Forecast annual revenue, 2026 and 2030. Source: Reuters / FT, September 18. Graphic: Superintelligence.
Anthropic’s IPO and OpenAI’s funding talks depend on investors backing future growth. Their leaders are also debating slower AI development. Today’s Vercel chart shows how quickly a new model can attract spending on one service. OpenAI’s projections show how much cash a lab can still need while sales are growing. I’d watch whether revenue can catch up with the cost of building and running the models.
Is Your Training Data Actually Model-Ready?
DNSMOS gives you a score, not whether that data fits your model. Treat it as pass/fail and you'll train on audio that looks clean but hurts performance, while tossing good data for no reason. Voices' CTO DJ Jalali just published a white paper with the four-step framework the team uses to set internal thresholds instead.
