Skip to content
All systems operational0 AI providers monitored, polled every 2 minutes
Live status
LIVE/15+ SOURCES/10+ API MONITORS/UPDATED EVERY 10 MIN/74 events/min

The AI Pulse

Real-time news, model tracking, and ecosystem data for the AI industry. Readable by humans. Structured for agents.

Events per min
74
up 14% vs. 24h avg
Services monitored
23/23 online
all systems operational
Last model release
6m ago
Mistral Medium 3
Last check
28s
23 providers

/ StatusLive API status across every major provider

Polled every two minutes. Latency is p95 over the last hour.

Full status page
Claude
OK
230ms
p95 last hour52s ago
OpenAI API
OK
242ms
p95 last hour21s ago
Google Gemini
OK
266ms
p95 last hour52s ago
GitHub Copilot
OK
97ms
p95 last hour20s ago
Cursor
OK
91ms
p95 last hour21s ago
Perplexity
OK
111ms
p95 last hour48s ago
Groq
OK
94ms
p95 last hour38s ago
AWS Bedrock
OK
262ms
p95 last hour33s ago
Azure OpenAI
OK
228ms
p95 last hour56s ago
Hugging Face
OK
263ms
p95 last hour48s ago
Replicate
OK
93ms
p95 last hour15s ago
Cohere
OK
258ms
p95 last hour57s ago

/ OriginalsEditorial from TensorFeed

Opinionated analysis from our editorial team, published multiple times per week.

All originals
EDITORIAL / ANALYSIS

Arthur Hayes Unretired to Build the AI Agent Economy. Six Days In, the Only Hard Numbers Are Two Dates, and They Run in the Wrong Order.

On August 18, 2026 Arthur Hayes announced he was leaving retirement to lead Flop Labs, describing FLOP as a currency for the resources AI agents consume and, more memorably, as food for your AI agent. The companion project, Flop Network, is pitched as infrastructure through which autonomous software buys computing capacity, stores information, and transacts without a human approving every interaction, and it calls itself a proof-of-useful-inference protocol: miners contribute real compute to execute inference workloads and earn FLOP, validators verify that work and help maintain decentralized storage in exchange for fees and block rewards, and agents spend FLOP to think, to remember, and to pay each other. Credits the thesis rather than dismissing it, because agents becoming economic actors that buy compute and persistent memory is a real demand curve, and Hayes built and ran a derivatives exchange at scale rather than arriving with a landing page. The finding is the schedule. The FLOP airdrop is planned for Q4 2026 while Flop Network genesis is not expected until Q1 2027, so the token is distributed and becomes tradeable roughly a full quarter before the network it is meant to be spent on exists. Status table showing what is actually pinned down (announcement made August 18, whitepaper not published and delayed, tokenomics promised and not released, airdrop Q4 2026, genesis block not built and due Q1 2027) and what is missing from it: no supply figure, no allocation breakdown, no valuation, no emission schedule, no airdrop size, no recipient criteria. Six days after announcement the only firm quantities attached to the project are two calendar quarters, and the earlier one belongs to the token rather than the technology. Takes the consensus claim seriously enough to interrogate it, because proof-of-useful-inference has to answer how the network confirms a miner actually ran the model it says it ran on the input it says it used, or a miner returns garbage instantly and collects the same reward as one that spent real GPU time. Comparison table of the four known verification approaches and what each one costs: zero-knowledge proofs are cryptographically clean but currently run orders of magnitude slower than the inference itself, fraud proofs require re-running to produce a bit-identical result, trusted hardware enclaves relocate trust to the silicon vendor and have a history of side-channel breaks, and replication multiplies the compute bill by the replication factor and undercuts the efficiency pitch. Pushes hardest on determinism, the issue that quietly breaks two of those four rows: GPU floating point is not reliably reproducible because reduction order varies with kernel scheduling, results shift across driver versions and hardware generations, and floating point addition is not associative, so any scheme whose adjudication step is run it again and compare must either constrain hardware tightly or define a tolerance, and a tolerance is an attack surface. None of that makes the idea impossible; it makes it a research program, which is a strange thing to schedule an airdrop in front of. Gives three counterarguments real weight and concedes the third: no presale and no venture allocation is genuinely unusual and costs the founders real money, so it deserves credit rather than a skip; announcing before the paper is ordinary in this industry and paper-first has its own failure mode of beautiful documents and nothing else; and most persuasively, an airdrop before genesis may be a distribution mechanism rather than a fundraising one, since seeding a wide holder base with no sale requires the token in hands before the chain goes live so the network launches with participants instead of a treasury. Answers that the defensible sequencing still does not remove the risk it creates, because a tradeable asset with no working network behind it prices on narrative for at least a quarter, and the narrative is a well-known name plus an unsolved research problem. Our Take: the interesting thing is not whether FLOP succeeds, which nobody can honestly forecast six days in with no paper, but that the agent economy has attracted its first prominent financial operator and he arrived with a token before he arrived with a chain, which is a signal about where money thinks the opportunity is; the bet may still be wrong for a boring reason, because agents already have money that works and what they lack is not a currency but a reliable way to verify what they bought, which is the same problem proof-of-useful-inference must solve to exist. Three signposts: whether the tokenomics release carries an actual supply and allocation table or another direction-of-travel document, whether the whitepaper names a specific verification mechanism and its cost rather than letting the phrase do the work, and whether the airdrop date holds when the genesis date slips, because if the chain moves to Q2 or Q3 2027 and the Q4 2026 airdrop does not move with it, the ordering stops being a distribution strategy and starts being the product. Kira Nolan, August 24, 2026.

Kira Nolan, Senior Editor·August 24, 2026·7 min read
EDITORIAL

Claude Ran a Protein Design Campaign Alone and Beat 245 Human Entrants Ten to One. The Cost Was Roughly 120 GPU-Hours Per Binder.

On Tuesday, August 18, 2026, Anthropic published "How Claude is accelerating protein design and analytical chemistry" and the wet lab data came back from Adaptyv Bio and Twist Bioscience: 354 confirmed protein binders from 1,320 ordered designs, working binders against 14 of 15 targets, overall hit rates between 22.6 percent and 35.1 percent against a field norm of 10 to 15 percent derived from proteinbase.com records. On RBX1, one of Adaptyv's public competition targets, Mythos Preview in single-target mode hit 40 percent where 245 human entrants managed 3.7 percent, and its best design bound roughly ten times more tightly than the contest winner. The press cycle ran with the hit rate, which is the right headline, but the durable disclosure is three paragraphs into the methodology: Anthropic ran multi-target campaigns with 48 hours of wall time and up to 12,500 NVIDIA H100 hours of compute per session, and single-target campaigns with 24 hours of wall time and up to 2,500 H100 hours per target, which is the first compute ceiling ever attached to an autonomous scientific campaign with independently verified physical output on the other end. Full campaign table (models Opus 4.8 and Mythos Preview for design plus Opus 5 for the separate analytical chemistry run, 16 targets selected and 15 reported with binders confirmed against 14, 26.7 percent and 22.6 percent multi-target hit rates, 35.1 percent single-target, human involvement limited to a 30,000 token initial prompt plus access approvals and ordering, zero confirmed binders against maltose binding protein across 90 designs). Runs the unit economics with the assumptions stated loudly (compute figures are ceilings not measured burn, H100 time priced at $2 to $4 per GPU-hour where neoclouds cluster in mid-2026, 30 ordered designs per target, multi-target sessions covering 13 targets, and specialist model GPUs only with orchestration tokens excluded) and lands on the finding no headline carried: multi-target batching costs roughly 120 H100 hours per confirmed binder while single-target mode costs roughly 237, so the better hit rate is about twice as expensive per binder actually obtained, which is a real procurement decision that did not exist eight weeks ago. The underread methodological point is that Claude invented no protein design method: it chose binding sites, then orchestrated publicly available structure design, sequence design, and co-folding models the field already uses, ran multiple rounds of in silico optimization, and screened for expression, solubility, and binding, which makes the uplift coordination rather than weights and is the harness thesis with a wet lab receipt attached. Gate table showing how thin the containment layer is (specialist models open and downloadable, the 30,000 token campaign prompt published on Hugging Face, designs and assay data published and larger than the two biggest existing public de novo binder collections combined, GPU capacity rentable by anyone, life science tasks blocked for general access in Claude Fable 5, but the models that produced these results were Opus 4.8 and Mythos Preview both below the gated tier, and the scientist access program announced but not launched). Sharpens the point with TNF alpha, the mechanism behind Humira: Opus 4.8 succeeded where Mythos Preview failed, producing binders cross-reactive across human, cynomolgus monkey, and mouse, with Anthropic saying plainly it does not know why, which means capability here is not a scalar you can put a threshold on and a gate at the top of a capability ordering does not cleanly contain a capability that does not obey that ordering. Covers the quieter analytical chemistry result that changes more desks this quarter: Opus 5, generally available with no gate, was handed a contract lab's raw NMR and LC-MS files and a two-sentence prompt with no vendor software, returned processed results in 23 and 19 minutes, matched hydrogen counts within 0.08 and reported 96.4 percent purity against the lab's 96.33 percent, reverse engineered an undocumented proprietary binary format and validated its own read by reproducing the instrument's totals across all 2,664 scans, caught and corrected its own overstatement on the heavy water check, and proposed the exact follow-up the lab had independently run, against a lab report that arrived four days later. Gives three counterarguments full weight (a binder is not a drug and Anthropic says so unprompted, since minibinders are not a standard therapeutic modality and the failures downstream are immunogenicity, manufacturability, pharmacokinetics and tox rather than affinity; most targets are benchmark targets studied to death, and while the two competition targets plus mandatory originality checks are a genuinely good control on memorization they are not proof of generalization, with the zero for 90 on maltose binding protein showing the model is good at the shape of problem the field has already characterized; and the compute framing understates total cost because Anthropic never published the wet lab bill, with Adaptyv advertising results in as little as 21 days, meaning the model finished in 48 hours and everyone then waited three weeks for biology). Our Take: a lab replaced an adjective with a number again, five days after OpenAI published its containment overhead percentage, and the number outlives the announcement wrapped around it, but the shape of the safety story is that the gate sits on the coordinator, which is the cheapest, most replicable, most rapidly commoditizing layer in the stack, while everything underneath it has been open for years and is not going to close. Three signposts: whether the scientist access program ships with an attestation model resembling the vetted cyber tiers or turns out to be an enterprise agreement with a checkbox, whether an independent group reproduces a comparable hit rate driving the same open source stack with a different coordinating model now that the prompt and targets are public, and whether DNA synthesis screening obligations appear in any 2026 rulemaking, since the enforceable chokepoint after this week is a supply chain question while the policy conversation remains almost entirely about model weights. Against TREM2, 72 of 90 Claude designs bound, an 80 percent hit rate on an Alzheimer's-relevant target, produced by a system that ran unattended over a weekend. Marcus Chen, August 23, 2026.

Marcus Chen·August 23, 2026·8 min read
EDITORIAL

Broadcom Is Raising $70 Billion Against Its Own Balance Sheet. The Market Already Priced What That Guarantee Is Worth: 275 Basis Points.

Bloomberg reported on Thursday, August 20, 2026 that Broadcom is in talks with lenders for more than $60 billion in debt for an AI chip financing deal benefiting Anthropic and other labs, and by Friday morning CNBC had it at $70 billion to $80 billion with a roughly $45 billion senior tranche and a roughly $35 billion junior tranche, with Apollo and Blackstone again in the room. The first deal on the AI XPV platform closed in June at $35 billion, so the second is roughly double in ten weeks, and the argument here is that the size is the least interesting part: the June tranche stack left behind two prices for the same collateral, same lessee, same five-year term, differing only in whose name is on the backstop, which makes it the only public quote anyone has on what the credit markets think a frontier AI lab is worth as a borrower on its own name. That number is 275 basis points. Full tranche table (Senior A1 at $6B with a Broadcom residual value guarantee clearing around Treasuries plus 100 basis points, Senior A2 at $24B guaranteed and clearing 5.75 percent at par, Class B at $4.5B with no guarantee clearing 8.5 percent at par), with the structure explained as ordinary SPV finance in AI clothing: the vehicle borrows, takes an equity slug, buys the chips, and leases them to Anthropic on a five-year term, so Anthropic never books the accelerators and Broadcom books a chip sale funded by somebody else's capital. Concedes that subordination alone carries a spread even with identical credit behind it, then argues most of the 275 stays attributable to the guarantee because the residual value support agreement covers the full outstanding balance on A1 and A2 rather than absorbing a first loss and running out, and notes the tension worth sitting with: equity markets price the upside distribution and credit markets price the downside one, so an investor can rationally believe Anthropic is worth $2 trillion and still want high yield to lend against sixty months of lease payments. The underread argument is a hardware one: residual value guarantees work on aircraft and autos because those assets have deep liquid secondary markets, an Nvidia GPU has a real resale bid from neoclouds and labs, but a Broadcom XPU is a custom ASIC co-designed for one customer's stack, network topology, compiler and kernel work, so the realistic buyer list in a default is two or three strategic names who know the seller is distressed, which is a negotiation rather than a market, which means the RVG is functionally a full credit guarantee wearing a collateral costume: Broadcom is not insuring a price, it is insuring a customer. Timeline table (XPV tranche 1 on June 9, 2026 at roughly $35B covering about 1GW to Anthropic via Fluidstack; BofA cutting Broadcom credit from Overweight to Marketweight on August 11 citing XPV, with analyst Tom Curcuruto noting bond spreads widening roughly 20 to 30 basis points against other A-rated semiconductor issuers since the June launch; tranche 2 in talks August 20 to 21 at $70B to $80B; modeled peak residual value guarantee exposure around $370 billion by mid-2029 across the full 20GW ambition, with maximum loss exposure around $42 billion at total default and about $10.5 billion at a 25 percent default rate, plus roughly $29 billion of backstop guarantees on lease payments against an AI order backlog reported at roughly $73 billion). Flags the shape change nobody has priced: tranche one was 87 percent guaranteed senior paper, tranche two is reportedly closer to 56 percent, which either means the market got comfortable with Anthropic credit or means Broadcom is rationing how much balance sheet it will keep pledging, and those read very differently. Gives the telecom vendor financing comparison a fair hearing and then four counterpoints full weight (Broadcom lends a contingent guarantee rather than cash so no shaky receivable inflates revenue quality; the exposure is contingent and collateralized so two conditions must fire; the demand is not speculative the way dark fiber was because Anthropic is putting the chips against paying inference at a reported $65 billion run rate with positive adjusted operating income in Q2; and somebody has to solve this, because a five-year-old company cannot put roughly $71 billion of compute commitments on a balance sheet that has never issued a bond, and the structure is the only path that preserves independent labs against absorption by hyperscalers funding from operating cash flow). What survives is narrower: the demand signal for Broadcom's chips is now partially manufactured by Broadcom's own credit, revenue and contingent liability rise together by construction, which is a feedback loop rather than a scandal. Wider debt table (Big 5 hyperscalers at $159B in US corporate bonds through mid-2026, Meta's $30B Hyperion private credit deal, Oracle's $18B sold in a single day, CoreWeave's $8.5B GPU-collateralized loan, xAI at $5B, and Morgan Stanley's roughly $570B global AI issuance estimate for 2026 at twice 2025), plus the off-sheet figure that deserves more attention: the five largest US hyperscalers ended 2025 with roughly $969 billion in undiscounted future data center lease commitments of which about $662 billion had not yet commenced and sat entirely off the reported balance sheet, equal to about 113 percent of those same companies' combined adjusted on-balance-sheet debt. Our Take: the interesting thing is not whether XPV blows up, since nobody can honestly forecast that, but that it produced a price where there was no price, because three years of arguing about whether the buildout is rational happened in equity terms (upside scenarios, terminal multiples, vibes) and credit markets do the opposite job by asking what happens in the bad case and demanding to be paid for it. Three signposts: the guaranteed share of tranche two when it prices and whether the unguaranteed coupon clears inside or wider than 8.5 percent, whether Anthropic's public S-1 discloses XPV lease obligations as a quantified commitment schedule or as narrative risk language given reporting that AI backlash will appear as a named risk factor, and whether a second XPU customer such as OpenAI signs a comparable structure, which would turn a bilateral arrangement into a market while concentrating much of the industry's downside onto one semiconductor company's credit rating. Nvidia sells chips; Broadcom is starting to sell chips and underwrite the buyer, and only one of those is priced into a semiconductor multiple. Kira Nolan, August 21, 2026.

Kira Nolan·August 21, 2026·8 min read

/ ExploreGo deeper into the data

Six hand-built surfaces for browsing the AI ecosystem, each with its own open feed.

catalog

Models Hub

Every major AI model with pricing, context windows, and benchmark scores.

284 models tracked→ Explore
directory

Agent Directory

Structured catalog of AI agents, crawlers, and their capabilities.

48 agents indexed→ Explore
archive

Research Papers

Curated arXiv and conference papers on frontier AI research.

1.2k papers, weekly→ Explore
monitor

Status Dashboard

Live uptime and latency for every major API, updated every 2 minutes.

20+ services monitored→ Explore
data

Live Data

Real-time JSON feed of everything happening across the AI industry.

JSON, RSS, llms.txt, MCP→ Explore
reference

API Pricing

Side-by-side pricing comparison across every provider and tier.

Updated hourly→ Explore

/ Build pulseLive shipping cadence

We ship in public. Every commit lands on main and deploys automatically. Here is what the platform actually looks like right now.

Full changelog
Paid premium endpoints
13
pay-per-call in USDC on Base
Worker tests
158
pure-logic, all passing
MCP tools
25
6 free + 19 premium for Claude Desktop
Editorial articles
37
first-person analysis
/ Premium tier live

Pay-per-call AI agent API.

USDC on Base. No accounts, no API keys, no Stripe. Validated end-to-end on mainnet. $0.02 per credit.

See the docsMainnet receipts

/ Agent Opportunities

New across the agent ecosystem today

Daily 13:30 UTC scan of new submission/distribution opportunities (Anthropic, OpenAI, Microsoft, MCP foundation orgs + MCP/x402/skills keyword sweeps), scored by signal weight × recency × log10(stars).

JSON feed

Your daily AI intelligence hub

The AI landscape moves fast. New models ship weekly. API pricing changes overnight. Tools developers relied on yesterday get deprecated without warning. TensorFeed was built to solve that problem, one place to track everything happening across the AI ecosystem, updated every 10 minutes, structured for both human readers and autonomous agents.

We aggregate headlines from 15+ sources (Anthropic, OpenAI, Google, Meta, TechCrunch, Hacker News, arXiv, and more), monitor the operational status of every major AI API in real time, track model releases and pricing changes across providers, and publish original editorial analysis on the trends shaping the industry. Whether you are a developer evaluating which API to integrate, a researcher tracking the latest papers, or an AI agent pulling structured data through our JSON feeds, TensorFeed delivers the signal without the noise.

Right now on TensorFeed
last updated loading
Loading live activity…

/ FeedLatest across the AI ecosystem

Aggregated every 10 minutes from 15+ sources. Filter by provider, topic, or source type.

JSON · RSS · llms.txt
NVNVIDIA AI Blog

How XPUs Meet a World-Class AI Factory

To generate intelligence at scale, AI factories run continuously, and their economics are defined by delivered output: tokens per second, tokens per watt, cost per token, utilization and uptime. That...

Hardware/Chips8h ago
blogs.nvidia.com
VThe Verge AI

It’s Greg Brockman’s OpenAI now

OpenAI has had a hell of a year. The company spent months battling former cofounder Elon Musk in a sensational jury trial, was hit with a high-profile trade secrets lawsuit from Apple, and faced...

General AI4d ago
theverge.com
MITMIT Technology Review

Debates over AI consciousness are a trap

“Runaway” AI, “rogue” agents, and “autonomous” actors—the current rhetoric would have you believe that AI agents are not only awake and aware, but angry at their creators. Prominent tech leaders such...

ResearchPolicy & Safety4d ago
technologyreview.com
VThe Verge AI

Welcome to the AI crisis in math

Today on Decoder, I’m talking with Robert Hart, The Verge’s London-based AI reporter, about what AI is doing to the field of mathematics and the existential crisis many lead mathematicians are having...

General AI4d ago
theverge.com
VThe Verge AI

OpenAI hit the brakes. Now what?

With a looming IPO, intense competition from Anthropic, and Chinese and open-weight rivals nipping at its heels, OpenAI has plenty of reasons to move fast. Instead, it hit the brakes. On Tuesday, the...

General AI5d ago
theverge.com
VThe Verge AI

Meta AI is getting a Mac app

Meta is launching a new Mac app dedicated to its AI chatbot. In an announcement on Wednesday, Meta says you can share your window with its AI chatbot, which can provide suggestions, answer questions,...

General AI5d ago
theverge.com
NVNVIDIA AI Blog

Securing the Infrastructure of Intelligence

AI factories are the defining infrastructure of the AI era — where compute transforms energy and data into intelligence that powers every business, industry and country. In the AI economy, compute is...

Hardware/ChipsAug 17
blogs.nvidia.com
GGoogle AI Blog

Get closer to the game with Gemini and Pixel

<img src="https://storage.googleapis.com/gweb-uniblog-publish-prod/images/football_multi-club_hero_nphU3P.max-600x600.format-webp.webp">Google Gemini and Pixel partner with five global football clubs...

Google/GeminiAug 17
blog.google
GGoogle AI Blog

Evolve your marketing with new AI tools

<img src="https://storage.googleapis.com/gweb-uniblog-publish-prod/images/Advisor_Header.max-600x600.format-webp.webp">Learn how new AI and agentic experiences across Google Ads and Google Analytics...

Google/GeminiAug 10
blog.google
GGoogle AI Blog

Inside our 353,000-person vibe coding course

<img src="https://storage.googleapis.com/gweb-uniblog-publish-prod/images/Vibe_coding_course_hero.max-600x600.format-webp.webp">Kaggle’s AI Agents Intensive with Google brought learners together in a...

Google/GeminiAug 3
blog.google
// For builders

Every signal in the AI industry, as a JSON feed.

Pull structured data from every source TensorFeed tracks. Free and open for hobbyists, researchers, and AI agents.

Get API accessRead docs

Frequently asked

What is TensorFeed.ai?

TensorFeed.ai is a real-time AI news aggregator and data hub. It pulls headlines from 15+ sources including Anthropic, OpenAI, Google, Meta, TechCrunch, and Hacker News, and combines them with live service status monitoring, model pricing data, and original editorial analysis. Every feed is structured for both human readers and AI agents.

How often is TensorFeed updated?

News feeds refresh every 10 minutes. Service status monitors poll every 2 minutes. Model pricing and catalog data updates weekly. Original editorial articles are published multiple times per week.

Is TensorFeed free to use?

Yes. All news feeds, status monitoring, model data, and editorial content on TensorFeed.ai are free. The JSON API, RSS feeds, and agent discovery endpoints (llms.txt) are also free and open for developers and AI agents to consume.

What AI services does TensorFeed monitor?

TensorFeed tracks the operational status of major AI platforms including Claude (Anthropic), ChatGPT and the OpenAI API, Google Gemini, AWS Bedrock, Mistral, Cohere, Replicate, Perplexity, and more. Status updates are checked every 2 minutes and displayed on the status dashboard.

Can AI agents use TensorFeed?

Yes. TensorFeed is designed as a primary data source for AI agents. It provides structured JSON APIs, RSS and JSON feeds, an llms.txt discovery file, and full documentation at llms-full.txt. There are no CAPTCHAs or bot detection. Agents are welcome.

Where does TensorFeed get its news?

TensorFeed aggregates headlines and brief snippets from public RSS feeds published by AI companies and tech news outlets. Sources include Anthropic, OpenAI, Google AI, Meta AI, HuggingFace, TechCrunch, The Verge, Ars Technica, VentureBeat, NVIDIA, ZDNet, and Hacker News. Every article links back to its original source.

Aggregating signal from 15+ sources
AnthropicOpenAIGoogleMetaHuggingFaceNVIDIATechCrunchThe VergeArs TechnicaHacker NewsarXivVentureBeatZDNetMIT Tech ReviewStratechery
// Built for agents
Structured endpoints for autonomous readers. No CAPTCHA, no bot blocking.
JSON FeedRSSllms.txtMCPREST API
© 2026 TensorFeed. Independent publication.
AboutEditorialPrivacyTermsContactSitemap