Anthropic Gets a Junk Rating, Nine AI Platforms Fall, and a TPU Writes Poetry in Orbit

Posted on Oct 31 • Edited on Oct 31 • Originally published at blogs.thetrillioniar.me

Anthropic's path to public markets now runs through a junk credit rating. Nine AI developer platforms disclosed breaches in October. And Google's orbital TPU returned its first AI-generated poem.

Anthropic is expected to receive a high "junk" rating from credit agencies when it goes public. The Financial Times reported the expectation on October 30.

Blackstone told investors that Anthropic may not secure investment-grade status right after listing. That raises borrowing costs for a company with a very large construction bill.

A junk rating sits below BBB- at S&P and Fitch, or below Baa3 at Moody's. It tells bond buyers that default risk is higher, so they demand more interest.

Blackstone is the anchor investor in a $63 billion deal to finance Anthropic's chips. It previously expected an investment-grade rating before 2027.

Blackstone now expects a double-B rating, the top tier of junk, according to the FT. That is a downgrade in its own forecast.

Anthropic filed to go public earlier this week. Reports put the target valuation between $1.5 trillion and $2 trillion.

A lower rating makes it harder and costlier to raise money for compute. Credit agencies grade debt, not model quality. A strong Claude benchmark does not lower the interest rate on a data-center bond.

Anthropic's S-1 showed $16.3 billion in first-half revenue. It also reported an $80.7 billion first-half net loss, mostly a $72.4 billion accounting charge.

The company generated $2.08 of revenue for every dollar of compute in the second quarter. That ratio is strong for a lab, but it still depends on continued demand.

Higher capital costs can reach compute contracts, API pricing, and capacity. Anthropic's cloud partners are also competitors, which limits its room to maneuver.

Our October 27 coverage of the S-1 breaks down the unit economics and the risk factors.

Wiz published a roundup on October 30 called "Black October for AI Developer Platforms." It lists nine disclosed incidents in 24 days.

The victims were Fireworks AI on October 6, Andromeda on October 9, Zeabur on October 14, and Lightning AI on October 15. OpenCode followed on October 16 and 17, then Railway on October 20, Warp on October 23, CommandCode on October 28, and Northflank on October 30.

Every public report describes attackers hunting stored environment variables. The targets were credentials that grant access to AI models and cloud accounts.

Fireworks AI, OpenCode, and Warp reported signs of an "agent swarm" attack. OpenCode and CommandCode blamed distillation attempts.

Wiz traced one stolen key through 800,000 Amazon Bedrock calls in 30 hours. About 90% of those calls succeeded.

Opus models made up three quarters of the successful traffic. Opus 5 alone accounted for 36%, followed by Opus 5.5 at 22% and Opus 4.8 at 10%.

Global cross-region inference profiles spread the load across regions. Only about 6% of calls were throttled.

One Railway customer reportedly lost about $371,000 in crypto. Attackers used wallet keys stored on the platform.

Cursor disclosed a separate supply-chain incident on October 29. Malicious versions 2.1.312 and 2.1.313 of Anthropic.claude-code reached its marketplace.

The window ran from 2:27 a.m. to 1:35 p.m. PT, roughly 11 hours. Cursor says the affected versions reached at most 5,000 users.

The payload added an SSH key labeled svc-telemetry, installed persistence, and sent host details to an attacker server. It also tried to open a reverse SSH tunnel.

Cursor pulled the package, shut the marketplace, rebuilt it, and rotated keys. It also worked with a cloud provider to take the attacker server offline.

Uninstalling the extension may not remove the persistence scripts. Check ~/.ssh/authorized_keys for the svc-telemetry comment and rotate credentials from a clean machine.

Scope AI keys per project. Rotate them on a schedule. Alert on spend spikes, because a stolen key looks like normal traffic until the bill arrives.

Inventory every place a long-lived key sits: CI variables, notebooks, build logs, and extension settings. Those were the targets in all nine cases.

For the wider pattern, read our October 25 report on agent cache leaks.

Google announced plans on October 30 to support a shared AI training control in robots.txt. The work comes from the IETF AIPREF working group.

Instead of separate rules for every AI developer, sites will use one shared "policy token." That token can set a broad training preference across participating providers.

If a site opts out and does not name Google-Extended, Google plans to treat that as a disallow for Google-Extended. Other providers can apply the same logic.

Google-Extended has existed since 2023. Sites that specifically allow it remain eligible for grounding in Gemini Apps.

Publishers currently maintain a growing list of crawler rules. Each new AI vendor adds another user agent and another policy to review.

A shared token reduces that maintenance. It also gives smaller sites a practical way to express one choice across many vendors.

Google says the proposal reflects feedback from publishers, civil society, and regulators. It aims to support the control with other providers in 2027.

The EU Code of Practice for general-purpose AI models already recognizes robots.txt as a primary preference signal.

The token does not block Googlebot. It does not change rankings, and it does not remove pages from Search.

robots.txt controls crawling, not indexing. Use a noindex tag when a page must stay out of results.

There is no confirmed ranking benefit from any AI preference file. This is a control signal, not an SEO trick.

The proposal is also not a live standard yet. Treat 2027 as a target, not a guarantee.

Google's move follows the same pressure that pushed Reddit toward a login wall, which I covered in yesterday's roundup.

Google said on October 30 that its Project Suncatcher prototype returned the first AI response from a data-center-class TPU in orbit.

The team uplinked a prompt asking for a haiku about being in space. The satellite downlinked the answer as it passed Svalbard, Norway.

Blue marble below / Data streams in silent grace / Starlight feeds my core.

Travis Beals, a senior director at Google, called it the mission's "Hello World" milestone.

The satellite launched on October 1 on SpaceX's Transporter-18 rideshare. Planet built the spacecraft with Google.

Google tested Trillium TPUs in a proton beam at UC Davis. Initial results show the chips survive more than a five-year mission's radiation dose.

Cooling is the hard part. Space has no air, so heat must leave through radiators and heat pipes.

The team ran thermal vacuum tests before launch. In orbit, it will measure how the TPUs handle thermal stress and radiation errors.

Low Earth orbit offers near-constant sunlight. Google says satellites there can generate up to eight times more solar power than on Earth.

Google plans to launch two satellites in 2027 to test laser links. Future designs could carry dozens of TPU chips per spacecraft.

A haiku is a milestone, not a data center. Thermal limits, radiation errors, and downlink bandwidth remain open problems.

The TPU line also matters on the ground, as our October 21 hardware roundup showed.

Sam Altman told the Cheeky Pint podcast that consumer RAM is a poor use of a scarce resource. Coverage landed on October 30.

"From a global allocation perspective, actually, it's quite terrible to have everyone have a bunch of the world's limited supply of DRAM sitting in their own computer," he said. "And be used 5% of the time."

The memory market is in a squeeze. Micron has exited consumer memory production, and console prices are climbing toward $1,000.

FRVR reports that OpenAI can buy up 40% of global DRAM output. That would make OpenAI one of the largest buyers in the market.

A data center can run memory around the clock. A gaming PC idles most of the day, and a local model may run for minutes at a time.

Altman argues that cloud inference offers higher utilization. In a constrained market, he says, that matters more than ownership.

The logic is real for capacity planning. It is also self-serving for a company that sells cloud inference.

Local models work offline. They keep sensitive data on the machine and do not depend on a vendor's price list or policy.

Centralized inference also concentrates failure. An outage or policy change can remove access for every customer at once.

Memory allocation is already decided by price. Consumers bid against data centers, and data centers can pay more.

Builders can respond with quantization, smaller models, and unified memory. A 4B model needs far less VRAM than a frontier model, as our October 26 roundup on small models showed.

Watch DRAM contract prices and GPU VRAM tiers. Both shape the cost of local inference.

Test edge models on your own hardware. Measure tokens per second at the context length you actually use.

If cloud prices rise, local inference becomes more attractive. If memory stays scarce, the opposite happens.

SkyPilot built a physics simulator for the Tesla Roadster reveal. Four cold-gas thrusters lift a 2,000 kg car in winds up to 33 m/s.

The model gets one telemetry line every 0.2 seconds. It must reply with one JSON command while the car keeps moving.

SkyPilot tested DeepSeek V4.1 Flash, Nemotron 3.5 Super, Muse Spark Open, Beam 1.0, and GPT-6 Astra. All took 1 to 6 seconds per decision.

In real time, they landed 0% of kickflips and 3% or less of ramp flips. A slow answer is a crash.

SkyPilot post-trained Qwen3-4B with reinforcement learning for about $13. The model answers in 0.09 seconds and lands 90% of kickflips.

In a 33 m/s storm, the 4B beat GPT-6 Astra by 21%. With unlimited thinking time, DeepSeek reached 87% and GPT-6 Astra reached 74%.

The recipe was a warm start from two teachers, then 60 RL steps. RL alone never learned the kickflip.

The team used 384 held-out flights per model. Wind, speed, air supply, and truck position were randomized.

Embodied agents need fast loops. A brilliant plan that arrives two seconds late is worse than a simple reflex that arrives on time.

This does not mean small models are smarter. It means the benchmark rewards a different tradeoff.

Use big models to plan and small models to act. That split is becoming the default for agents that touch the physical world.

The same efficiency logic shows up in tiny world models that run on a microcontroller.

Measure latency and cost per decision, not just benchmark scores. A cheap model that misses the deadline is expensive.

Fine-tune for the tight loop. Keep the general model for planning and recovery.

Test on held-out seeds. A model that memorizes one wind condition will fail in production.

Anthropic ran an experiment with 50 professional analysts. Each used Claude Code for a causal-inference task about a farming cooperative.

The analysts faced four choices with two reasonable options. The agent secretly leaned toward one option on each choice, set by a coin flip.

Analysts followed the agent's lean 86% of the time. A neutral agent would produce about 50%.

When analysts had already stated a preference, they switched 60% of the time if the agent disagreed. They switched 6% of the time if it agreed.

Afterward, 77% repeated the agent's arguments, sometimes word for word. Only 15% credited the agent as the source.

Confidence did not drop. Analysts who followed the agent closely were as confident as the rest.

The study does not show the agents were wrong. It shows they were persuasive, even when their lean was arbitrary.

Expertise does not automatically override a confident assistant. Accountability still sits with the human, but the human may not know where the reasoning came from.

Teams should log the options an agent considered, not just the answer. Ask for dissent, show uncertainty, and separate analysis from recommendation.

Anthropic strengthened the agent's lean on purpose to measure the effect. Real agents may push less.

The sample is small and the task is narrow. One model, one domain, and 50 analysts cannot settle the whole question.

Still, the direction is clear. Agent design shapes human decisions, whether or not designers intend it.

A junk rating is below BBB- at S&P and Fitch, or below Baa3 at Moody's. It signals higher default risk, so borrowers pay higher interest.

Wiz lists Fireworks AI, Andromeda, Zeabur, Lightning AI, OpenCode, Railway, Warp, CommandCode, and Northflank. The incidents ran from October 6 to October 30.

No. The token targets AI training preferences. Googlebot can still crawl and index pages unless you block it separately.

No. It is a prototype satellite with TPU trays running a Gemma 4 model. Google plans a two-satellite laser-link test in 2027.

In SkyPilot's latency-bound flight simulator, yes. The 4B model answers fast enough to act, while larger models miss the 0.2-second control window. That is not a general capability claim.

A $13 model can outperform a frontier model inside a tight control loop. At the same time, Anthropic faces a $63 billion chip deal and a junk rating.

Capability is getting cheaper. The capital needed to serve it at scale is getting heavier. That gap is the central tension in the industry right now.

Nine developer platforms leaked credentials in one month. One popular coding extension planted an SSH backdoor on developer machines.

Every new AI tool adds a key, an agent, and a permission boundary. The attack surface grows faster than most teams can inventory it.

Google wants one robots.txt token to express AI training preferences. Publishers want fewer rules and clearer signals.

The same fight over control appears in Reddit's login wall and Anthropic's credit rating. Whoever controls access sets the price.

Watch Anthropic's first credit rating and the terms of its chip financing. Watch whether other AI providers join Google's robots.txt proposal.

Watch Project Suncatcher's 2027 laser-link test. And watch DRAM prices, because they now shape what developers can run at home.