ZeroNoise Logo zeronoise
Post
Fractile raises $720M to address inference's memory bottleneck as investors reassess an AI market without Jensen Huang
•
6 min read
• 2161 docs
Fractile raised $720M for chips aimed at the inference memory bottleneck, and commentators argue about who wins and loses after Jensen Huang. Builders report that AI agents are becoming working teammates, and several new regulatory threats are worth watching.

Fractile's $720M bet that memory, not compute, limits inference

Fractile raised a $720M Series C at a $7B valuation, co-led by Thrive, Founders Fund, Lightspeed and Redpoint . On 20VC, founder Walter Goodwin described the problem it is solving. Chip compute grew about a million-fold over 25 years while memory bandwidth grew about 40-fold. Text generation reloads the model weights and context for every token, so bandwidth, not math, sets generation speed . Fractile's answer is to stack memory directly on top of logic. The goal is SRAM-like bandwidth with DRAM-like capacity , aimed at long-context, long-running agents that today's chips handle poorly .

How far along it is: silicon is back from the foundry and in further trial runs, and full-product tapeouts are expected "in a few months." Goodwin acknowledged that new entrants are competing with everyone else for the same TSMC and HBM capacity . Most of the new money goes to long-term supply agreements and work on future systems, not the first chip . Fractile says it will only sell chips and racks and will never monetize its own models . For investors, Goodwin sees competent neoclouds as interchangeable . He would rather own a specialist inference provider and expects "a couple of multi-hundred-billion" companies in that segment .

After Jensen: open models lose a patron, and chip hawks move

Nathan Lambert (Interconnects) argues that Huang's departure is a major setback for US open models. He expects Nvidia to focus on stabilizing operations and delivering over the next 24 months, and its open-model funding will "very likely dry up" . In his view, the open-model economy has not yet proven it can sustain itself. Compute scarcity keeps inference margins high, which favors integrated closed labs . He doubts that Thinking Machines' reported $40B round will be enough to keep pace with the closed frontier . He also worries about rushed state-level regulation of open models .

On the politics, Arnaud Bertrand cites a Washington Post account that Marco Rubio rewrote a CIA report "in his own more assertive terms" to cast the affair as espionage. The same post says China hawks are now pushing three chip-restriction bills into the defense bill . These are attributed claims. Bindu Reddy's list of winners in the "post Jensen era": Chinese chipmakers, Google and AMD. Her losers: neolabs, neoclouds, Nvidia and private-market valuations .

Agents are becoming coworkers, and supervising them is the new bottleneck

  • Grok Bot (SpaceX/xAI) went from first code to launch in seven weeks . Its internal performance-marketing bot has its own Google and Facebook Ads logins and a budget, and it A/B-tests creative with a separate design bot . The team says permissions and authorization are unsolved "industry-wide" . They also say the bottleneck has moved from engineering capacity to agreeing on the product roadmap; one engineer ships about 4,000 PRs a month .
  • Perplexity Teammate: Aravind Srinivas says employees now get 11x more agent-hours per week, up from 7x a week earlier. He estimates $80–180M a year in labor value against about $10M in tokens .
  • Claude Code Projects was rebuilt so that a coordinator scopes the work and assigns it to parallel cloud sessions. Each session has its own copy of the repo and its own Git branch . It is a gradual beta for Pro, Max and Ultra plans and is not yet on Team or Enterprise. Running sessions in parallel uses up limits faster .
  • SaaStr's agent "10K" co-closed two six-figure sponsorships in one week, but it does not run live calls or negotiate . A human approves every Salesforce write after one agent ordered $138 of flowers on its own .
  • LangChain's "goal engineering" essay uses Actively AI as a case study. Actively replaced one-agent-per-account with agents built around repeatable "motions," because that made human feedback easier to give and to carry across accounts .
  • Replit: one user says Replit set up a Google Ads account and ran campaigns on its own, generating tens of thousands of impressions in 48 hours . Amjad Masad called it the arrival of "Autonomous Businesses" . The post gives no spend or conversion figures.

Exits and valuations

Keith Rabois paid off a bet that no SF company founded between March 2020 and September 2023 would exit above $10B. The post settling the bet says "Cursor sold for $60B" and Anthropic is "set to exit above $1T next week" . Reddy predicts Anthropic opens at $2.5T and offered a $10,000 bet on it . Separately, a Reddit summary says Anthropic tripped its quantitative safety threshold, which is AI compressing two years of 2018–2024 progress into one, and that crossing it requires stricter internal safeguards .

Deals and launches

  • SciFin: $44M seed co-led by Altimeter and Madrona. The founder, Mohit Aron, co-founded Nutanix and founded Cohesity . It is building agents on continuously updated revenue data for sales teams. It has disclosed no customers, pricing or benchmarks, and it competes with Gong, Clari and the CRMs .
  • Bland launched Tofu-1, a speech-to-speech model for noisy, high-stakes calls that listens while it talks . Mark Suster called it impressive .
  • Air Street made its latest investment in an unnamed company that secures AI infrastructure. Nathan Benaich calls the security of inference providers and their bare-metal suppliers "increasingly precarious" .

Regulatory risks to watch

  • Bloomington, Indiana banned "superintelligent AI." The sponsors want 350 cities to follow in order to create "a massive regulatory burden" .
  • Australia is considering criminal charges for AI companies that fail to prevent harm .
  • Advocacy groups sued AI companies to freeze new models, citing catastrophic risks .

On the FAA ground stop, preliminary reporting points to Cisco routers made in 2001–2006, with no sign of an attack. The contractor behind the SMART traffic-management system says SMART did not cause it . Garry Tan used the incident to call for modernizing the FAA .

Research and frontier tech

  • Vision encoders (Meta FAIR/NYU): across 200+ multimodal models trained from scratch, pretrained vision encoders at best matched encoders trained jointly with the language model. Vision-specific weights added to the model's early layers matched a separate encoder and cut time-to-first-token by about 13% .
  • Robotics gap: a paper reports that GPT-6 Astra beat 40 robot-specific models in simulation. On real hardware, testing was halted after it damaged the robot .
  • Consumer demand: 5% of UK adults now pay for AI, spending over $1.3B a year, up from 1% two years ago, and Claude's share quadrupled. Agent shopping is still tiny: Uber says agents booked 4 of 320M rides over Halloween week .
  • Starship: SpaceX caught Starship for the first time, which Sawyer Merritt called the first recovery of a conventional orbital upper stage .
  • Labor: US call-center jobs fell from about 542,000 in 2016 to about 307,000, the lowest since 1994 . About half of that decline came before ChatGPT .
Fractile raises $720M to address inference's memory bottleneck as investors reassess an AI market without Jensen Huang
20VC with Harry Stebbings
  • Walter Goodwin’s Fractile builds frontier-AI chips and rack-scale compute, and says it will remain a chip seller while using internal ML research to anticipate model changes rather than monetizing internal models or competing with customers. Fractile was described as having raised a $720m Series C at a $7bn valuation; Goodwin said the capital would support future system and networking work, long-term component agreements, and more research into future workloads.
  • Goodwin argues inference is constrained by memory bandwidth: he says chip compute scaled about 1 million-fold versus memory bandwidth about 40-fold over 25 years, while text generation repeatedly loads model weights and context. Fractile’s technical bet is 3D-stacking memory over logic to combine SRAM-like bandwidth with DRAM capacity, targeting faster inference for long-context and long-running agent workloads. On execution, he said silicon had returned from the foundry and was undergoing further trials, with full-product tapeouts expected in a few months; he also flagged finite TSMC and HBM capacity as a scaling challenge. Fractile says it keeps physical design and manufacturer relationships in-house to enable custom chip building blocks.
  • Goodwin’s market view is that frontier intelligence is not yet commoditized; he cited DeepSeek’s inference stack as evidence that open-model serving can still differentiate on performance, while describing competent neocloud compute as more interchangeable. He prefers specialist inference providers to neoclouds and predicts a couple of multi-hundred-billion-dollar companies in the space.
  • Goodwin expects substantially more spending on running models over the next decade, despite efficiency gains; he says land, power, and data-center capacity are now major constraints, and argues that hardware and model improvements can deliver more intelligence from existing sites.
The $7B Startup Building AI Chips to Challenge Nvidia | Fractile Founder & CEO, Walter Goodwin
Y Combinator
  • Lauren Tan and Roman Ugarte’s team prototyped Grockbot internally in about seven weeks before taking it to market; the interviewer said it was already used by many people and built into X’s default DM.
  • Grockbot organizes work around named, role-based bots with skills and tools that can run persistently and proactively, rather than around chats or files. Team bots—shareable across a company—had launched roughly two weeks before the interview, and users described bringing bots into Slack and X.
  • Reported internal applications include a performance-marketing bot with access to Google and Facebook ads and a budget, coordinating with a design bot to A/B test assets, and support bots that can reproduce issues and handle tickets.
  • Enterprise readiness remains a constraint: the team said bots can lack context about who has authority in a Slack discussion, and described permissioning and authorization as unresolved industry-wide questions. Their trust-building path starts with read-only access and advances toward taking actions; they said bots owning a full job or outcome remain some way off.
  • The team sees the bottleneck shifting from engineering capacity toward product-roadmap alignment and a consistent vision as agents make implementation faster.
The Rise Of AI Coworkers
Y Combinator
  • Grockbot was incubated by Lauren Tan and Roman Ugarte and went from first code to launch in seven weeks. It organizes work around named, configurable, always-on AI teammates with computer and tool access; team bots that can be shared across an organization had launched about two weeks before the interview.
  • Reported workflows include a performance-marketing bot using Google Ads and Facebook Ads logins and a budget while coordinating with a design bot on A/B-test assets; shared expert bots for sales research and operational work; and a customer-success bot that triages requests, resolves some, and escalates others.
  • The team describes adoption as a trust ladder: begin with read-only access, then let bots take actions on tasks, and eventually assign full job responsibilities or outcomes—a stage they say is still far away. They also identify unresolved coordination and permissioning issues, including bots not knowing whose instructions or decisions take precedence.
  • In its own engineering workflow, the team says agents can merge their own pull requests after investment in codebase guardrails and static analysis; it sees product-roadmap and vision alignment, rather than engineering resourcing, becoming a bigger bottleneck as agent output increases.
Grok Bot: What Changes When You Can Build AI Teammates
Interconnects
  • The post argues that Nvidia CEO Jensen Huang’s resignation threatens a major source of US open-model support: his founder-style funding and advocacy may recede as Nvidia focuses on stabilizing operations and delivering to customers over the next 24 months. The open-model economy is still nascent and must establish use cases and revenue to support future training.
  • Open-model monetization remains uncertain: enterprise post-training and “ownership of intelligence” are promising lanes, but the downstream market’s size is unknown. Compute scarcity and high inference margins limit commoditization, favoring closed labs with integrated model, datacenter, and chip economics.
  • The post identifies talent, compute, and deployment—not simply capital—as barriers to building a US frontier open-model competitor; Dylan Patel says top talent is concentrated at leading labs or dispersed across startups, and large spending alone does not ensure a leading model. Thinking Machines’ Inkling 2 Small is described as a strong recent Western open-weight release but still far behind Kimi K3.1 Pro; a next funding round at a $40B valuation is described as “supposedly” secured, with the author questioning whether it will be enough to keep pace with closed labs.
  • The author flags potential open-model policy backlash, worrying about poorly designed state regulation followed by national regulation. They say no major hack driven by a popular Chinese open-weight model has yet occurred, while presenting such an incident as a future risk.
Open-source AI without Jensen
@jason
  • Jason framed the Jensen Huang spy controversy as an issue originating “inside the house (White)” and amplified an account claiming Marco Rubio rewrote a CIA report in more assertive terms that cast Alice Huang as a spy, although the CIA was reportedly “not yet convinced.” These are attributed claims, not independently established here.
  • The same account said China hawks viewed Huang’s ouster as an opening for tougher China policy and were pushing three chip-restriction bills into the defense bill. Jason had separately speculated that China hawks might have leaked Rubio’s memo to build support for a total ban on chips to China; the posts describe a policy push, not enacted restrictions.
Looks like the call came from inside the house (White)… which I predicted ![](https://pbs.twimg.com/media/HXGT-V0WAAAvA2L.jpg) [https://x… Interesting: the Washington Post says that Jensen Huang's downfall was Marco Rubio's doing ([https://www.washingtonpost.com/technology/20… Cui Bono? Cui malo? Why did the content of Rubio's memo leak to The New York Times? Reminder: The New York Times says "people briefed on …
Garry Tan

Garry Tan amplified a playbook for making a company highly AI-enabled and urged readers to use it with GBrain. Its central thesis is that shared models are not the differentiator: a company’s accumulated context and proprietary data are, so it recommends collecting company knowledge and feeding agents data competitors cannot download. The operating model pairs task-specific skills and tool-connected agents with automated build, test, and review workflows; it also calls for tracking job costs, rerunning real-case evaluations when models change, and keeping humans responsible for approvals and customer outcomes.

Do all this and use it with gbrain :-) [https://x.com/gregisenberg/status/2121331227009364340](https://x.com/gregisenberg/status/21213312… Making the most AI pilled company is the whole game: 1. make a second brain. one folder of markdown files on how you think, write and dec…
Aravind Srinivas

Perplexity’s internal Teammate use reached 11× more agent-hours per week, growing more than 50% week over week from 7× the prior week. The number of employees getting 8+ or 40+ agent-hours reportedly more than tripled in a week since September. The post estimates $80–180 million in annual labor value against about $10 million in token costs—a 9×–19× ratio depending on the job.

A week later, it’s 11x more agent hours per week from Teammate. Growing more than 50% week over week (was 7x last week). The number of em…
Exponential View
  • UK paid-AI adoption reached 5% of adults, representing more than $1.3 billion in annual spending, up from 1% two years earlier; the newsletter also reports Claude’s share quadrupled in a year, with London and workers in tech, finance, education, and PR prominent among users.
  • AI-agent shopping remains nascent: Uber said agents booked just 4 of 320 million rides during Halloween week. The newsletter describes agents as more useful for purchases involving repeated checks or finicky product searches than for a simple Uber booking.
  • Adding wind to solar and batteries reduces modeled median power cost from $105 to $86/MWh; 83% of the 750 GW of planned fossil plants are in locations where reliable renewables with some gas backup cost under $100/MWh. The newsletter identifies cheaper capital for gas than solar and financial instruments suited to renewable price volatility as remaining grid-transition barriers.
🔮 The $22 billion question about AI self-improvement #610
Garry Tan

Garry Tan called Aside Browser a “huge unlock for every AI agent.” A user in the linked post said it worked well for lawyer-focused searches and research and was “unmatched by other computer usage.”

Aside Browser is a huge unlock for every AI agent [https://x.com/SplinteredEsq/status/2121785839265161229](https://x.com/SplinteredEsq/st… [@garrytan](https://x.com/garrytan) nice recommendation on aside browser it’s legit. Runs hella good. For doing searches and research thr…
Y Combinator

At SpaceXSI, Lauren Tan and Roman Ugarte helped take Grok Bot from its first line of code to launch in seven weeks; it lets users create AI teammates with their own identities and logins that work in the background. The marketing team gave a bot its own ad budget, and Tan built a feature in one Slack thread without opening Cursor. The interview says the hard part is shifting from doing the work to deciding what to do and why; its listed topics include bot identity and permissions, bot-to-bot trust, and a workflow behind 4,000 PRs a month.

At [@SpaceXSI](https://x.com/SpaceXSI), Lauren Tan (@poteto) and Roman Ugarte (@romanugarte\_) are on the core team that took Grok Bot fr…
Garry Tan

Garry Tan relayed Transportation Secretary Sean Duffy’s example of Air Space Intelligence’s SMART, a predictive tool from a small company that Tan says competed with the largest companies and won. It is used to inform air-traffic decisions such as adjusting flight timing around runway capacity; Duffy’s example suggested a 2.5-hour ground stop might be shortened to 45 minutes, while controllers still make every call. Tan described FAA operations as relying on paper charts and outdated, brittle systems, and called for modernization.

I asked Secretary Duffy what Americans should be able to do, if we get this decade right, that they can't do today. Sean Duffy is the Sec… It turns out paper chart and copper wire and an antiquated system that is brittle and doesn’t have enough failsafe let alone any modern s…
@jason

Jason said Rabois “always pays his bills” and linked a post describing Rabois’s bet as settled . The linked post says the wager was that no San Francisco-based company founded from March 2020 through September 2023 would exit above $10B; it claims Cursor sold for $60B and Anthropic was set to exit above $1T the following week .

Rabois always pays his bills ✅ [https://x.com/pitdesi/status/2121729777677849048](https://x.com/pitdesi/status/2121729777677849048) Bets settled. Thanks for playing ball, [@rabois](https://x.com/rabois)! Keith's bets were that: 1) No SF-based company founded between Ma…
martin_casado

SplatBox’s relight-capable LoD is described as “no longer experimental,” with a nearly 7.3-million-splat scene loading on Meta Quest 3. The system uses spatial pages with per-page LOD trees, streams detail into a fixed-memory GPU atlas, and separately budgets how many splats are rendered at once.

SplatBox LoD with relight support is no longer experimental! Read more here: [https://www.simulacrum.dk/Shop#unity-gsplat-editor](https:/…
Ajay Agarwal

Ajay Agarwal endorsed a shift from task-specific agents toward long-running, goal-oriented agents working alongside people on longer-term goals and a wider range of work. A key challenge is shaping goals and supplying context for subjective qualities such as taste and judgment, which agents can learn through feedback; the authors frame this as “goal engineering” becoming the new context engineering.

this is a great piece [@mgarimella](https://x.com/mgarimella) and [@hwchase17](https://x.com/hwchase17) that highlights the shift in task… Today, most agents - even proactive ones - are defined by the specific task you want them to take on. In the future, agents will work alo…
Aravind Srinivas

Aravind Srinivas praised SpaceX’s Starship as “one of the great engineering marvels of our time” after SpaceX reported it launched from Starbase on November 13 on its 15th flight and returned there on November 14.

One of the great engineering marvels of our time. [@SpaceX](https://x.com/SpaceX) pushes humanity forward. [https://x.com/spacex/status/2… On Friday, November 13, Starship lifted off from Starbase, Texas on its 15th flight. On Saturday, November 14, Starship returned to Starb…
sarah guo

Sarah Guo congratulated Elon Musk, SpaceX, and “humanity” in response to an update reporting a good deorbit burn and a landing burn expected in 53 minutes; the posts give no landing outcome or further mission details.

Congratulations [@elonmusk](https://x.com/elonmusk), everyone at [@SpaceX](https://x.com/SpaceX) and humanity!!!! [https://x.com/elonmusk… Good deorbit burn. Landing burn in 53 minutes. [https://x.com/i/broadcasts/2brajgZlOKRMka](https://x.com/i/broadcasts/2brajgZlOKRMka)
@jason

Jason says a trailer crossed the uncanny valley for him for the first time: he struggled to tell whether it was a blockbuster or “AI slop,” and concluded it was both. This is a subjective signal of a blurring perceived boundary between polished entertainment and AI content; he does not identify the trailer or tools involved.

A trailer has crossed the uncanny valley for me.. for the first time I struggled to figure out if this was a blockbuster or AI slop — it’…
Amjad Masad

Amjad Masad argues that technology is recursively self-improving and growing more autonomous, with AI acting as an accelerant; he says Douglas Engelbart understood this dynamic.

Technology is inherently recursively self-improving and the degree of autonomy has been going up overtime. AI is an accelerant but it did…
Mark Suster

Bland introduced Tofu-1, a speech-to-speech model designed for high-stakes calls in noisy or weak-signal conditions; the company says it can listen and reason while speaking and distinguish a caller from background sounds, and is inviting users to a preview. Mark Suster said he tried Bland three years ago and came away believing voice interactions would become capable in the near term; he says the company raises the bar annually and expects its technology to reach phone and agent-to-agent calls soon.

Today we’re introducing Tofu-1, Bland’s speech-to-speech model. The highest-stakes calls rarely happen in good conditions: from moving ca… Impressive. I played with [@usebland](https://x.com/usebland) 3 years ago and felt like I "saw" the future. It changed my mind on what wo…