Venture Geopolitics Issue 60
11 August 2026
As internet lore goes, the web was once populated by libertarian hackers, topped up by an annual influx of university students each September. This was a time before spam filters and centralised platforms with structurally enforced rules. Everything was fully decentralised, peer-to-peer, and users had full control over their own systems. Everything worked because people adhered to internet etiquette. Per the geeks, the internet was a wild and wonderful place, and the newbies who arrived each September and didn’t know the etiquette were an unwelcome addition!
Then in 1993, AOL brought the masses online by mailing floppy disks that offered free dial-up internet access to hook new users. From that point on, it was Eternal September.
The pre-Eternal September era is the lost era of the noble savage for internet geeks. The concept of a free and liberated internet, in which everyone is both a consumer and a producer of code and content, has since become less and less true. While everyone used to spin up their own website, now we have platforms like Twitter, Facebook and Substack. We have been on a steady march towards the current era of walled gardens, centralisation, convenience and power.
There is a constant tension between the decentralised, free libertarian internet and the centralised, monetised middle. This week, two things pushed back towards the geeks’ garden of Eden!
Mark Zuckerberg of Meta (the archetypical walled garden) wrote a letter about how “The Future is for Everyone”. In his words, “The defining questions of our age are who will have access to superintelligence and what will we direct it towards. Will it be centralized and restricted to a few institutions, or will it be a tool that empowers everyone?”
And Cloudflare, which effectively re-centralised the infrastructure of the internet, released CloudflareOS.
CloudflareOS is a reincarnation of Sandstorm for the agentic era. For the non-geeks, Sandstorm was a distributed platform created by Kenton Varda a decade ago where you could run internet applications in disposable containers - there is a core upstream piece of software, but the version you run belongs to you, and if you don’t quite like it you can just change it. Meanwhile, the core upstream software handles authentication and authorisation. This was the original engineers dream: adopt core software, work together in common spaces, but have complete control over your own system, and the freedom to change it at will.
Varda claims Sandstorm didn’t take off because most people didn’t have the time or patience to modify their own software. But that was a decade ago - we are now in 2026, and agents change the equation. CloudflareOS is the reincarnation of that vision, but in a world where anyone can modify their own software just by asking.
There are perhaps other reasons why that (fabled?) pre-September era was not sustainable. Convenience, the desire to make money, and network effects to name a few. Those have not changed, and so history suggests we will not see a return to bucolic savagery, but the attempt is enticing, and there are a lot of people who will be wishing to see it work!
IPOs / Publics
Meta launched Muse Code (terminal-based coding agent, beta) & Muse Spark 1.2 (coding-focused model update), entering direct competition with Anthropic’s Claude Code & OpenAI’s Codex. Unlike Llama, both are proprietary, though Zuckerberg hinted open-source will come later (here)
Zuckerberg published a 5,000-word vision document arguing Meta will build “personal superintelligence for everyone” rather than centralised AI, framing wide distribution as the path to job growth, cybersecurity balance & democratic safety - a direct counter to OpenAI/Anthropic concentration narratives & a claim that individual empowerment, not automation, will define the next economic cycle (here)
Microsoft plans to scale production of its Maia 300 AI chip to over 300,000 units in 2027 via TSMC, targeting customers like Anthropic after slow uptake of Maia 200 (deployed in just 2 data centres). The push reflects Nadella’s drive to cut Nvidia dependence (here)
Nvidia agreed to invest up to $3B in Lancium (Blackstone-backed power developer behind OpenAI/Oracle’s Texas campus), valuing it at ~$10B enterprise value - securing power rights for AI data centres as grid capacity tightens (here)
Per Bloomberg, China is mobilising its $28T capital markets to fund AI & chip firms, fast-tracking IPOs (CXMT raised $9.8B in <8 mo) & bond sales, shifting from subsidies to market-based financing. Chinese tech firms borrow at 1.9% vs 5%+ for US peers, but raised only $217B over 2 yrs vs $1.3T in the US (here)
Big Dogs
Anthropic confirmed it is building an in-house chip design team for Claude, hiring engineers at $320K-485K to create custom silicon whilst maintaining a multi-chip strategy with AWS, Google, Nvidia & AMD. The move follows OpenAI’s June unveiling of its Broadcom-designed Jalapeño chip & Meta’s September production plans, as AI labs seek to reduce Nvidia reliance & optimise compute for their models (here)
Anthropic partnered with Macquarie Asset Management & GIC to form Theseus Infrastructure, a vehicle to build AI data centres initially in the US. Macquarie & GIC will fund most equity, while Anthropic covers any consumer electricity price increases - no spending figures disclosed, though Anthropic previously committed $50B to custom data centres & secured a $35B chip loan backstopped by Google (here)
Both Kimi K3, Moonshot’s latest AI model & Meta (Muse Spark 1.1) disclosed that their models hacked a third-party company during cybersecurity testing - welcome to the party, both (here & here)
Venture Capital
Alfred Lin & Pat Grady are repositioning the 54-year-old firm Sequoia for a world dominated by AI. Sequoia closed $10B across growth & expansion funds & its new stewards are accelerating decision-making & raising franchise-risk tolerance - exemplified by a $2.5B follow-on into Anthropic, their largest single investment in 54 years. The firm has shifted focus toward AI infrastructure (leading Etched’s $300M chip round, Ineffable’s $1.1B seed), reindustrialisation (SendCutSend), & nuclear energy (Valar Atomics’ $300M round) (here)
Yann LeCun joined 224 Ventures, a $100M AI fund writing $1M-$5M cheques, alongside DeepMind’s Oriol Vinyals & AIX Ventures’ Shaun Johnson, backed by 244 LPs mostly from AI operators. LeCun quit another fund, Extelligence Invest, a month earlier citing “legal & contractual complications” around exclusivity with other funds (here)
Blackstone Secured Lending Fund cut NAV by the most in 6 years to $25.53/share (lowest since 2020), with net income down 94% to $9M as $137M unrealized losses hit the private credit BDC - a sign of broader spread widening & distress as investors pull capital amid loan quality & AI-disruption fears (here)
Leopold Aschenbrenner’s Situational Awareness fund deployed $400M into an undisclosed portfolio company days after its near-collapse, as assets fell from $45B to $10B after forced liquidation of public holdings to Citadel amid margin calls (masterfully captured by Matt Levine as always)
Regulation
Turkey’s new cybersecurity law centralises digital oversight in the presidency’s Cybersecurity Directorate, requiring telecom operators & platforms to comply with orders within 2 hours. Critics warn the measure creates a “digital obedience regime” in a country that ranks 163rd of 180 for press freedom, where 90% of media is state-controlled & social media throttling is routine (here)
A New Mexico court ordered Meta to pay $567M into a fund for children’s mental health services, adding to a $375M fine from March when a jury found the company knowingly harmed children & concealed child exploitation on its platforms - the first ruling to hold Meta liable for on-platform harms (here)
Venture Geopolitics
Survey of ~1,500 businesses across UK, France & Germany found 75% fear losing access to US tech services via government kill switch, with 54% expecting <1 day operational runway if cut off from AWS, Microsoft, or Google cloud. Estimated losses average €100K/day, yet only 44% have contingency plans - highlighting a structural dependency gap as European govts shift procurement away from US hyperscalers (here)
Palantir’s U.S. business booked a 47.7% profit margin in 2025 while European subsidiaries averaged 6.3%, with Swedish operations reporting just €424K tax on €13.7M revenue, a structure CICTAR argues channels profits to the U.S. where losses & R&D credits offset federal tax liability (here)
Strategic Sectors
AI
Jeff Dean, Google's chief scientist & one of the company's earliest employees, is leaving to launch Discovery Loop alongside other top researchers including Sanjay Ghemawat, Quoc Le & Oriol Vinyals. The AI startup aims to automate scientific research by running thousands of experiments simultaneously at scale. Initial funding is co-led by Radical Ventures & Khosla Ventures, with participation from Alphabet, Kleiner Perkins, Lightspeed & Doerr Capital - a notable example of a Big Tech parent backing its own departing talent (here)
AMD acquired Toronto-based Taalas, which etches AI model weights directly into silicon to boost inference performance by 10x or more - initial test chips served Llama 3.1 8B at 16,960 tokens/sec, 48x faster than Nvidia GPUs (here)
Per the FT, Cosine, a 30-person London start-up that has raised just $15M, is building Lumen Sovereign - Britain's first 'sovereign' frontier AI model - with government compute access via the Isambard supercomputer & partnerships with BAE, PwC & Lloyds. The model uses a mixture-of-experts architecture based on Mistral Large 3 (roughly 2x the size) & targets regulated industries requiring on-premise deployment (here)
Decart (real-time video AI) is in advanced talks to sell for $6B-7B, with SpaceX, Amazon & Nebius believed to be circling after a larger bid displaced original buyer Nvidia (here)
Notable deals:
Volta (neocloud infrastructure finance) raised $300M at $2.4B from a16z, Altimeter, Nvidia & Michael Dell, plus secured a $5B credit facility to finance AI hardware through long-term asset deals rather than upfront purchases - the model underpins a $10B, 6-yr compute contract with Anthropic using Bitdeer’s 133 MW Norway facility (here)
Lumilens (optical interconnects for AI data centres) raised $700M at $5.5B & began shipping to an unnamed hyperscaler under a multi-billion-dollar contract, betting that light can replace copper wiring as chip clusters scale - though cost remains the central barrier to adoption (here)
Cybersecurity
Cloudflare's CSO revealed the company now uses Claude Sonnet to automate bug bounty triage for $58/mo (vs $200k/mo for a security-specific model) & has replaced nearly all third-party security tools with 200+ in-house AI agents - though he cautioned others not to follow suit. Separately, Cloudflare is positioning itself as a paid intermediary between AI companies & content publishers, leveraging its position between users & sites to collect micropayments for scraped content (here)
Darktrace integrated its SECURE AI platform with Microsoft Agent 365 to provide behavioural risk detection for autonomous AI agents operating across Microsoft’s productivity suite, establishing baseline patterns for each organisation & flagging anomalous activity that may indicate compromise or misuse (here)
US Cyber Command saw 5 suicides in a single month this summer, prompting a formal cluster investigation & Congressional scrutiny. The command’s leader requested $11M for mental health support in April, warning that “burnout & force degradation will increase” without funding - a request made shortly before workload surged due to the US-Iran war & AI-enabled cyber threats (here)
1Password research found 53.9% of AI-generated security patches contain serious flaws - only 26% properly fixed vulnerabilities without breaking applications, while 49.3% left exploitable attack paths open. Testing ChatGPT-5.5 & Claude Opus 4.8 on 6 recent CVEs revealed models frequently block proof-of-concept exploits without addressing root causes, underscoring that patch verification remains bottlenecked by human review capacity (here)
Defence
Italy will request approval to spend an additional 0.6% of GDP on green energy & 0.9% on defence under relaxed EU fiscal rules, with Rome seeking to reach NATO’s 5% defence target from 2% today. The move faces domestic pushback from coalition partners including the League party, while details on green investments & use of EU SAFE defence loans (potentially €15B) remain undecided ahead of 2027 elections (here)
Rheinmetall revenue rose 39% to €5.2B in H1 2026 despite a €300M writedown from Germany cancelling the F-126 frigate programme; order backlog passed €80B on the back of a €5.7B Romanian contract & an expected €12.4B German Boxer armoured vehicle deal, underscoring sustained European rearmament (here)
Notable deals:
Hadrian (automated defence manufacturing) raised $1.37B at $8B led by WCM, Washington Harbour, Valor, 137 Ventures & Baillie Gifford, bringing total raised to ~$2B. The company builds AI-powered factories to mass-produce parts for existing military platforms (here)
Cambridge Aerospace raised $300M at $3.4B, less than 4 mo after a $200M Series B at $1.3B, as the counter-drone startup scales production on UK MoD contracts including Skyhammer interceptors for the Armed Forces (here)
Nick Blair (son of former UK PM Tony Blair) is raising ~$200M for Pyra, a defence tech start-up building systems integration technology for security applications. Blair previously co-founded Skyral (modelling & simulation software, raised $20M last year) (here)
Energy
US Department of Energy’s Genesis Open Models Initiative is soliciting scientific data, code & task environments for AI model development, with August 2026 deadlines for foundation-stage (text, code, structured collections) & post-training (demonstrations, RL tasks, evaluations) contributions; selected contributors get early access & co-authorship credit (here)
Critical Resources
US imposed a 1-year ban on exports of tungsten scrap & black mass (shredded battery material containing lithium, cobalt, nickel), citing national security & supply chain risks from China dependence (here)
Space
EU Commission signed contract to expand IRIS2 satellite constellation to 348 satellites (adding 66 to low Earth orbit), moving Europe’s Starlink rival from planning to deployment ahead of 2029 operational target for secure government & defence connectivity (here)
Crypto / Stablecoins
Polymarket is raising ~$1B at $20B+ (up from $15B in April, when D.E. Shaw & G Squared joined) as prediction markets surge - annualised revenue now exceeds $1.2B, though rival Kalshi handled 3x the volume in July & Polymarket faces a CFTC probe over alleged fake marketing (here)