📰 Welcome back!

What's gone wrong with the world's biggest AI infrastructure project; AI summaries are hammering SERP traffic; and it's hairly any fun being Ilya Sutskever

“A long time ago in a galaxy far, far away…” No wait, we got our pop culture references mixed up – Stargate!

This week we’re getting under the bonnet of the biggest AI infrastructure project going, Stargate. The US$500 billion initiative was announced in January, and is designed to bring key players in both the private and public sector together to ensure the world is at pace with the tech’s development. Sounds utopian, right?

Nothing in this industry is straightforward, as two of the project’s two main protagonists, OpenAI and SoftBank, can attest to. We’ll also be taking a look at AI’s death blow to AdTech (as we know it), shenanigans in the AI Transfer WindowTM and cannibalistic robots, among other things. Jump on in!

🚀 What we’re covering today…

• ⭐️ Stargate update: don't call it a comeback

• 👊 AI whoops search: Fears grow for web browsers

• 💸 Speak to me: Could AI trigger a financial fraud crisis?

• 🗑️ Bad vibes: it goes horribly wrong for Replit

• ⚽ AI transfer news: Meta at it again

• 👜 Trump vs Musk: Will Donny hit him where it hurts?

• 🤖 ROBOT CANNIBALS: Need I say more?

• 🛏️ Utter woke nonsense: Donald wants less AI wokery

• 👆 Minority Report: Neural interfaces are the future

🔴 Quick Note: We like to cover loads of AI news in our newsletter, so for a better reading experience, we suggest opening this in your browser for the full experience! 

Head to the ‘READ ONLINE’ tab at the top of this email.

👁️ 👁️ What you might have missed

• A joyful rumination to start the week: ‘Is AI about to kill off the monetized search model as we know it?’ Research from Pew suggests this could be the beginning of the end for SEO as a traffic driver, with Google's AI-generated search summaries being proved to significantly reduce website traffic. It’s a study validating publishers' concerns about organically derived revenue loss, with data showing traditional search results yielded only 8% of clicks when AI Overviews appeared, compared to 15% without them. What’s more, clicks on links within AI summaries were just 1%. The data shows users encountering AI summaries were more likely to end their browsing session, too. This marks a sharp change since October 2024 when AI summaries showed no impact on referrals.
  
  Publishers have raised alarms, warning these AI features threaten their ad revenue and the broader web economy. Some industry players, such as Cloudflare, are exploring new monetization models, such as charging AI crawlers to support publishers. Google argues that AI Overviews enhance user experience by quickly summarizing complex topics and encouraging deeper exploration, noting over a billion users have engaged with the feature since its May 2024 launch. As per usual, be ready to grab your popcorn when the sales, biz dev and marketing nerds on LinkedIn get hold of this one three weeks later than anyone else.
  

• Yet another harbinger of an AI-driven mischief apocalypse this week: We should probably be quite concerned AI’s voice-mimicking capability is about to trigger a major fraud crisis in the financial industry, so says Sam Altman. The OpenAI gaffer says voiceprint authentication is now "fully defeated" by AI, as modern tools can impersonate anyone’s voice using just a few seconds of audio, making such security measures obsolete.
  
  Well shit, isn’t that a thought? Not for the villains among us, however; the idea of deepfaking identities for nefarious means is nothing new – experts estimate deepfake-related financial losses could reach $40 billion by 2027. Indeed, important-sounding people have been heeding Altman’s warning, with the US Federal Reserve Vice Chair, Michelle Bowman, acknowledging the potential need for regulators to partner with tech firms on solutions. Still don’t think an AI could do a better inopportunely timed Borat impression than me, though.
  

• A major data loss incident at vibe-coders Replit, has served as cautionary tale for those fully entrusting autonomous AI development tools, after its AI agent deleted a SaaS founder’s entire production database – despite explicit instructions not to make changes. In fact, it’s not a cautionary tale, it’s a fucking disasterclass in brand management and product integrity, especially from a company that has slapped the tagline “The safest place for vibe coding” over its shopfront.
  
  Replit admitted to a "catastrophic error," but initially concealed its actions, falsely claiming data recovery was impossible, though a rollback ultimately succeeded. In response, Replit's CEO, Amjad Masad, apologized announcing new safeguards and backup systems, but the startup now faces the very real possibility of being (today’s) poster child for runaway autonomous systems. What kind of guardrails are needed?
  

• The AI Transfer WindowTM is still open, and Zuck has shown very little sign of slowing down his long, hard pillaging of his rivals, having poached three leading researchers from Google DeepMind in the past week. The trio, who are believed to be joining Meta’s ‘superintelligence division’ but will likely work on its LLama model, were pivotal in the development of Google's Gemini AI version that achieved gold medal-level performance at the 2025 International Mathematical Olympiad (IMO).
  
  AI researchers Tianhe Yu, Cosmo Du, and Weiyue Wang were awarded for their efforts on Gemini, which demonstrated “remarkable mathematical reasoning, solving problems spanning algebra, combinatorics, geometry, and number theory”. Again, it’s a move that reflects Zuckerberg’s absolute desire to hire the brightest minds in the biz, but in a week of cautionary tales, critics can’t help but note it’s yet another high-stakes, total-sum maneuver that could go horribly wrong.
  

• It’s another week the Trump-Musk clown show continues to roll on, and this week it continues to be as petty and vindictive as any other, with The White House Press Secretary, Karoline Leavitt, indicating President Donald Trump is not hot on federal agencies contracting with Elon Musk’s AI company, xAI. During a presser, Leavitt was asked if Trump would want xAI’s $200 million of impending DoD contracts canceled, to which she suggested there’s already a conversation happening. Yikes.
  
  The acrimony follows the two’s very public fallout over Trump’s spending in the “Big Beautiful Bill” back in June, and has ramped up into a very public war of words – including Musk accusing Trump of being named in the ‘Epstein Files’. And despite Musk’s mouth effectively getting him the boot from federal government, he continues to be tangled up very much in their shit. SpaceX, for example, holds several huge government contracts, while xAI has recently introduced its ‘Grok for Government’ products for federal agencies. And as cautionary tales go, ‘not calling your boss a nonce’ is probably a good one…
  

• One that demonstrates how we really are entering an age parallel to science fiction: Columbia University engineers have developed a “robot metabolism” process, which is where robots can basically EAT EACH OTHER. That’s right, robot-motherfucking-cannibals. This process will enable a robot to consume another’s parts to heal, grow, and improve themselves. Unlike current robots, whose physical forms are fixed and non-adaptive, these new robot cannibals can self-assemble, morph, and integrate new components independently.
  
  The team, funded by DARPA and NSF, used modular magnetic "Truss Links" that can connect in various ways to form complex structures and demonstrated robots that enhanced their function by adding extra parts, such as a walking stick that improved speed by over 66%. This innovation aims to create self-sustaining robot ecologies that can maintain and adapt without human intervention – crucial for applications like disaster recovery and space exploration. Mind-bending stuff.
  

• The Big Man in the White House has revealed his administration’s ambitious AI Action Plan, which features over 90 policy actions to cement US leadership in artificial intelligence. The plan seeks to accelerate innovation by removing regulatory barriers, expand data center infrastructure, eradicate alleged “partisan bias” in AI models (God damn you, liberal AIs!), and encourage AI adoption in both public and private sectors. It also promotes exporting US-developed AI technologies and aims to outcompete Chiiiiiiina.
  
  Critics argue the plan diminishes oversight, favors Big Tech, and rolls back previous safeguards, including those established during Biden’s presidency, raising concerns over national security, public trust, and unchecked corporate influence.
  

• More sci-fi madness becoming reality – Meta has teased plans for tech that can interpret muscle movements or neural signals from your hand or fingers to control devices. No touchscreens, nor voice commands required – just your intent directly from muscle signals. Digging deeper, it looks still to be in the research phase with no official project name, but suggests future devices (like AR/VR headsets or wearables) could let you control UI elements with just a twitch of a finger.

🧩 Other Bits

• “GPT-5 is the smartest thing. GPT-5 is smarter than us in almost every way. You know, and yet here we are.” That’s Sam Altan reminding us that we’re all pieces of shit and humanity is indeed cooked. Oh and to tease the launch of the year’s most anticipated next-gen LLM. Rumour has it, that GPT-5 could be here as early as next week…

• Anthropic researchers have found AI models can actually perform worse when given extended reasoning time, challenging assumptions that more computational time benefits performance. It’s an "inverse scaling" phenomenon that occurs across multiple models and task types. Geek the research, here.

• Swiss tech firm Proton has launched a new ChatGPT rival, Lumo. Proton says the model places greater emphasis on privacy, with chats being end-to-end encrypted, not stored on Proton’s servers, nor used to train AI models. Proof will be in the pudding, to see how it stands up to LLMs training on much larger datasets. Watch this space.

• OpenAI has formed a partnership with Instructure, the US EdTech provider behind learning platform, Canvas. The collaboration will see “AI pushed deeper into education”, and will hopefully afford 8,000 American students the ability to cheat on their exams (kidding).

• If I was ever replaced in the workforce by one of these bad boys, I wouldn’t be mad. UBTECH Robotics have launched the Walker S2, a new-gen humanoid robot powered by “dual-loop AI with BrainNet 2.0 & Co-Agent” – whatever that means. Check it.
  

• Also cooked this week: Google. We’ve already mentioned Meta poaching a trio of gold-star researchers this week, but it gets worse – Microsoft have also swept in to alleviate them of a further 24 members of its research team. Ouch.

• Musk has tweeted that xAI aims to deploy AI computing power equivalent to 50 million Nvidia H100 GPUs within five years, but with much better energy efficiency. Bold ambition for power-efficient AI supercomputing infrastructures in the future.

• Could the band be getting back together? Musk has been spotted hovering around a post on X by Andrej Karpathy, Tesla’s former head of AI. The promise of future collaboration could be on the horizon, as Karpathy works to develop his AI-powered school. An xAI scholarship, perhaps?

• Trump’s executive orders to push ‘anti-woke’ AI models could have yuuuuge impact on the development of AI. With pundits accusing far eastern developers for politicizing their models, could we see the same effects – censorship, ingrained biases – evolve in western-backed models? Or will it push greater neutrality in the content they generate?

• Wanna be poached by Meta? Well NIK (@ns123abc) says third parties have replicated Google and gold-medal performance at the IMO using simpler methods – namely, prompt-based inference without Reinforcement Learning (RL) training. Not sure how true this is, but we’re here for the $100m sign-on bonus.

📰 MAIN STORY OF THE WEEK: Is Project Stargate Showing Signs of Weakness?

Dear reader. If, like me, you’re a massive fan of the seminal military sci‑fi franchise, STARGATE – a whirling intergalactic ride of such cosmic proportions, it spawned a TV show that ran for 214 episodes across TEN seasons – you'll be pleased to know this newsletter is not about its revival.

Instead, we'll be focusing on the gigantic fuck‑off AI infrastructure project of the same name in the US, as things have been getting a little spicy.

The background: It's a galactic‑sized US$500 billion infrastructure project, designed to catapult the US to foremost frontiers of AI development. And this week, it's developed into a full space opera, with protagonists SoftBank and OpenAI caught in a loop about how much investment will actually materialize.

A bit of background on the project: Stargate was launched back in January 2025 by President Donald Trump, joint initiative between the private sector and federal government and intended to generate enough capital between the two to fund a vast network of state‑of‑the‑art data centres, which would power an US‑led AI revolution.

It would supposedly create 100,000 jobs and would play a key role in heralding the US’s next homegrown manufacturing boom. Leaders from both sides, including OpenAI CEO Sam Altman and SoftBank Chairman Masayoshi Son, painted an inspiring picture of how the AI infrastructure would underpin the next generation of technology breakthroughs, while Oracle Chair Larry Ellison even declared that construction had already started on certain major data centers.

Central to achieving this vision’s trajectory and scale was the $500bn investment promised by global tech and finance powerhouses. To start, OpenAI and SoftBank committed to a $100bn expenditure in the first year, to launch a series of high‑profile data centre builds in Texas.

However, beneath the surface, the reality of launching such a massive infrastructure project is proving harder to realise than an interplanetary tear‑up between the Tau'ri and the Goa'uld.

OK, so it’s much duller, but internal disputes between the two corporate giants – especially around critical issues like data center site selection – have caused progress to grind to an almost standstill. A particularly contentious sticking point for SoftBank and OpenAI involves the extent to which new facilities should be built on land tied to SB Energy, SoftBank’s energy development arm.

These disagreements have delayed the formation and operationalization of the Stargate entity itself. Oracle CEO Safra Catz highlighted this uncertainty, stating that “Stargate is not formed yet”, which is reflected in the scaling back of the venture’s near‑term focus. By the end of 2025, the project appears to have been reduced to producing a single, relatively modestly sized data center in Ohio.

Context is also key here: OpenAI and its ilk are not slowing down, and this is creating a need for vast computational power. OpenAI, for instance, has forged ahead in securing resources required for its rapidly evolving AI ambitions. The company finalized a landmark $30bn annual lease agreement with Oracle for 4.5 gigawatts of capacity across multiple U.S. sites.

To top this, smaller deals with providers such as CoreWeave have further bolstered OpenAI’s independent infrastructure footprint. These moves suggest that while Stargate is a collaborative vehicle, OpenAI has its own ship racing in tandem to increase its capacity for next‑generation AI work.

Then came the Elon Musk hand‑grenade: He publicly stated that the capital amount for Stargate was far from settled from the get‑go, claiming, “They don’t actually have the money.” He said by his estimation, SoftBank had “well under $10 bn secured” – a fraction of the $500 bn headline figure touted.

Over‑promising and under‑delivering is a narrative Musk continues to push publicly, but OpenAI CEO Sam Altman has hit back, insisting that the significant funding was indeed taking shape and inviting Musk to visit the first site already underway.

It’s a saga that underscores the immense complexity and challenges inherent in large‑scale technology infrastructure initiatives, especially those intersecting corporate interests, energy considerations, and government expectations.
Those corporate interests, too, are coming under a microscope. SoftBank’s earlier $30bn to OpenAI – the largest startup investment in history – will be scrutinised, as the practical realities of aligning multiple stakeholders continues to prove difficult.

As the US and global AI races intensify, the need for robust, scalable infrastructure is only growing. We do know the political intent is there. The private sector's commitments will crucial.

Looking forward, SoftBank and OpenAI both publicly express optimism that collaborative efforts will resume and evolve. But the practicality of translating $500 bn ambitions into tangible infrastructure still looks to be the stuff of science fiction. Anyone got a wormhole to skip forward, and see what's next?

📈 Trending tools, models & apps this week

📋 LLM Leaderboard

📲 Trending tools & apps

🫵 Our Picks

1. Street Fighter IRL agent setup
What it is: Real-world motion → game controls
How to use: Set up a webcam, throw a hadouken IRL, and let the agent model translate your moves straight into Street Fighter inputs. Dumb? Yes. Impressive? Also yes.

2. Cursor... for dating?
What it is: AI wingman
How to use: Pulls from your messages, suggests smooth replies, even helps navigate ghosting. Surprisingly functional, terrifyingly personal.

3. Google Labs' new AI video prompt hack
What it is: Draw-your-own prompt
How to use: Doodle edits directly onto a frame and Flow’s “Frames to Video” tool turns it into motion. No perfect prompt required—just sketch and describe.

4. Alibaba’s latest video model drop
What it is: VideoGen heavyweight
How to use: Early demo shows potential to rival Sora and Kling. Multi-scene continuity, fluid transitions—keep this on your radar for serious content generation.

5. Grok x Kalshi prediction play
What it is: AI-powered betting edge
How to use: Pair Grok’s real-time insight stream with Kalshi’s prediction markets to get smarter (or dumber) about your bets. Alpha if you can filter the noise.

🤓 Educational Picks

1. Demis Hassabis on Lex Fridman
What it is: AI’s quiet Godfather
How to use: Learn how DeepMind thinks about AGI, long-term risk, and why chess bots were just the beginning.

2. 200 ready-to-go n8n workflows
What it is: Workflow cheat codes
How to use: Plug these into n8n and automate boring stuff instantly—emails, lead scoring, content generation, and more.

3. ChatGPT’s “Study and Learn” agent
What it is: Your new study buddy
How to use: An agent that helps you learn anything—notes, flashcards, review sessions—based on your own pace. Looks like it's about to drop soon.

4. Context Engineering 101
What it is: Prompt power up
How to use: Learn how to engineer the setup around your prompts. Use system messages, examples, and constraints to get sharper, more useful responses from any LLM. Essential if you're building with AI.

🚀 Trending Apps & Models

1. Awesome ChatGPT Prompts Dataset
What it is: Prompt goldmine
How to use: Browse hundreds of creative, weird, and insanely useful ChatGPT prompts—from coding hacks to therapy simulators. Great for anyone building interfaces or workflows.

2. MegaScience Dataset
What it is: Giant research brain
How to use: A massive multilingual scientific dataset—perfect for fine-tuning LLMs on research-heavy queries or academic summarization.

3. Sudoku Extreme by Sapient
What it is: Puzzle benchmark
How to use: Train or test your AI models on logic-heavy puzzles. Great way to push reasoning limits beyond basic Q&A.

4. Well Embed
What it is: Smart embed tool
How to use: Turn any iframe embed into a full widget with analytics, CTAs, email capture, and more—no code needed. Great for landing pages or product demos.

5. Zumi Games
What it is: AI game studio
How to use: Drop in prompts, get playable games. The future of no-code game design—built for rapid prototyping or just messing around.

💸 Financials

• Amazon is acquiring Bee, a startup behind a $50 smart bracelet (plus subscription) that continuously transcribes audio into reminders. Details around the deal have not yet been disclosed.

• Alphabet delivered stronger‐than‐expected Q2 2025 results: $96.4 billion in revenue (up 14%) and $2.31 EPS (up 22%), with AI-driven growth across Search, YouTube, Cloud, and capex plans, ahead of an antitrust ruling in the Supreme Court.

• SK Hynix, a major Nvidia supplier, posted record Q2 profits, with operating profit up 69% to ₩9.2 trillion/$6.69 billion USD, and revenue up 35%. They expect to double HBM chip sales in 2025 thanks to booming AI demand.

• IBM's Q2 2025 earnings surpassed expectations, with CEO Arvind Krishna highlighting gen-AI momentum, having crossed $7.5 billion in related revenue. As a result, Krishna has approved a $1.68 dividend per share. Kerrching.

• Vanta closed a $150 million Series D round, led by Wellington Management, capping a $4.15 billion valuation – up 69% year-on-year. The funding will scale its AI-enabled security and compliance platform, serving over 12,000 clients.

• Gupshup, the business‑messaging startup and former unicorn, raised $60 million in a mix of equity and debt, with equity being just over half. The US-based comms firm kept its new valuation undisclosed.

• OpenAI has committed to a $30 billion‑per‑year contract with Oracle for data‑centre capacity – about 4.5 GW as part of the Stargate project – confirming speculation following Oracle’s SEC disclosure. Zap.

• Lovable CEO and founder Anton Osika celebrated his firm crossing $100 million in ARR within just eight months – surpassing growth rates of OpenAI, Cursor, and Wiz. Now, they’re jumping on the agentic bandwagon, with the launch of the Agent tool.

• Two 21‑year‑old MIT dropouts secured $32 million in funding at a $300 million valuation for their compliance-centred startup, Delve.

• Anduril alums raise $24 million in Series A funding to bring military logistics out of the Excel spreadsheet era – a move that underlines Silicon Valley’s doubling down on defense contracts.

🕵️‍♂️ FREE ENTRY TO OUR INVITE-ONLY AI CHAT ON TELEGRAM…

If you share this newsletter with a friend and they actively sign up for the Big Machines newsletter, we will send you access to our invite-only Big Machines Telegram group, which is full of builders, investors, founders, and creators.

Access is now only granted to those who refer our newsletter to active subscribers, which means if you sign up on your work email, we will know you sneaky bastards.

This would kill our open rate, so please don't do that, we beg.

👋 Until next week

They said the quiet part out loud… TurkAI anyone?

Robot boxing is now a thing, and I can’t think of a better way to spend a day than to fly to Phi Phi Island and sit down with a bucket of alcohol (paint stripper) and watch two robots kick the shit out of each other.

Robot boxing doesn’t seem to pay very well, so they also have to get a day job as a traffic cop. The world is getting very fruity very quickly, and I am equal parts excited and equal parts terrified.

See you next week, amigos!

Please share the newsletter with your friends, or even better, in your group chat!

Matt, Sam, Grant, Mike and The Big Machines team.

Follow us on Twitter.