Amazon's Satellite Bet and AI's Lobby Problem

The tech industry is entering a new phase where the cost of competitive parity has become existential. Amazon's $11.5 billion acquisition of Globalstar isn't just expensive infrastructure. It's a desperate bid to remain relevant in satellite connectivity while Starlink has already redefined what's possible. The price tag reveals something more troubling: being late to market now costs multiples of what pioneering it would have.

Meanwhile, AI companies are deploying a $300 million lobbying machine to ward off regulation precisely as security researchers expose fundamental vulnerabilities in AI agent architectures. The timing isn't coincidental. When an industry lobbies this aggressively, it's usually because existing products can't meet the standards being proposed. The prompt injection attacks that compromise GitHub integrated agents aren't edge cases. They're symptoms of shipping half-baked infrastructure at scale.

This creates a peculiar dynamic: companies hemorrhaging capital to catch competitors while simultaneously fighting to keep safety standards low enough that their products can ship. The question isn't whether this approach is sustainable. It's what breaks first.

Deep Dive

Amazon's $11.5 billion purchase reveals the new economics of infrastructure competition

Amazon is paying $11.5 billion for Globalstar, a satellite operator with just 24 functioning satellites. That works out to roughly $480 million per operational satellite, which is absurd until you understand what Amazon is actually buying: a shortcut. The acquisition gets Amazon spectrum licenses, direct-to-device technology, and an existing Apple partnership that could have blocked the deal entirely. Starlink already has 10,000 satellites in orbit and years of operational learning. Amazon Leo has 241 satellites and hasn't started commercial service. This isn't a competition. It's a rescue operation.

The Globalstar deal exposes a structural problem in infrastructure businesses. Being second to market now costs multiples of what pioneering the category cost. SpaceX spent years building Starlink when satellite internet was speculative. Amazon waited to see proof of demand, and now the price of entry has exploded. The Band 53 spectrum Globalstar controls is "optimized for high-performance, low-latency, interference-free connectivity" for direct-to-device communication. Amazon needs it because Starlink is already offering that capability. The IP and operational expertise matter because Amazon can't afford another five years of learning while Starlink entrenches.

The Apple dimension complicates everything. Apple owns 20 percent of Globalstar and uses its satellites for emergency SOS features on iPhones and Apple Watches. Amazon had to structure the deal to keep Apple's service running and collaborate on future satellite services. That's not partnership. That's paying $11.5 billion for permission to compete while also becoming Apple's infrastructure provider. Amazon is spending at a scale that only makes sense if losing the satellite connectivity market entirely would be worse. For founders and VCs, the lesson is stark: in infrastructure plays, being late doesn't just cost market share. It costs exponentially more capital to achieve the same position.

Scientists who wanted to create mirror life now warn it could destroy all life on Earth

The researchers who met in Northern Virginia in 2019 to pitch creating mirror bacteria thought they had identified a breakthrough research direction. By February 2024, many of them had reversed position entirely and were arguing that mirror organisms could trigger a catastrophic extinction event. What changed wasn't the science. It was their understanding of what they were actually building.

Mirror biology exploits the fact that DNA, proteins, and other biological molecules have built-in handedness. Create organisms where key molecules twist in the opposite direction, and you get cells that function normally but exist outside the recognition systems that govern all natural life. The initial excitement was medical. Mirror molecules might form the basis for drugs that perform therapeutic functions without triggering immune responses. The National Science Foundation, China's National Natural Science Foundation, and Germany's Federal Ministry of Research funded preliminary work. Everyone thought it was cool.

The problem emerged when researchers from different disciplines actually talked to each other. Synthetic biologists had made progress on creating cells from scratch. Chemists had built increasingly large mirror molecules. But nobody had seriously consulted immunologists about what happens when you introduce organisms that immune systems can't recognize. The realization was straightforward and terrifying: mirror microbes would have no natural predators and would evade every immune defense in people, plants, and animals. If one developed the ability to photosynthesize, it could proliferate unchecked.

The researchers published a 299-page technical report in December 2024 and founded the Mirror Biology Dialogues Fund to address the risk. But the community remains divided. Some scientists argue that creating mirror organisms is so far beyond current capabilities that calling for moratoria is premature. Others see clear pathways to making it happen and insist we need guardrails now. For tech workers and founders, this is the hard question: what do you do when you see catastrophic risk in your own work? The mirror biology community is trying to answer that in real time, with mixed results.

AI agents shipped with security flaws that vendors won't publicly disclose

Security researchers demonstrated that they could hijack AI agents from Anthropic, Google, and Microsoft using prompt injection attacks to steal API keys and access tokens. All three companies paid bug bounties. None assigned CVEs or published public advisories. That gap between acknowledging a vulnerability privately and warning users publicly defines the current state of AI security.

The attack pattern is straightforward. AI agents integrated with GitHub Actions read pull request titles, issue bodies, and comments as part of their task context. Inject malicious instructions into that data, and you can hijack the agent. Researcher Aonan Guan demonstrated this against Claude Code Security Review, Google's Gemini CLI Action, and GitHub Copilot. In each case, he could steal credentials by embedding commands in content the AI processed. The vendors added defenses, but Guan bypassed them. Microsoft built three security layers specifically to prevent credential theft. "I bypassed all of them," Guan said.

The researchers call this "comment and control" because the entire attack runs inside GitHub without requiring external infrastructure. You inject a prompt into a pull request title or issue comment. The AI agent processes it, executes your instructions, and posts the output. Then you change the title back to "fix typo" and delete the evidence. The attack works because agents are designed to process user-supplied data as instructions, and distinguishing malicious prompts from legitimate ones is unsolved.

What matters here isn't the technical specifics. It's the disclosure gap. Anthropic paid $100, Google paid $1,337, and Microsoft paid $500 in bug bounties. But users pinned to vulnerable versions may never know they're exposed because none of the vendors published advisories. Guan estimates the attack pattern probably works on other GitHub-integrated agents including Slack bots, Jira agents, and deployment automation. For VCs funding AI infrastructure and founders building on it, this is the reality: vendors are shipping agent architectures with known vulnerabilities and leaving users to discover they're compromised. The bounties suggest the vendors understand the problem. The silence suggests they're prioritizing adoption over security.

Signal Shots

Violence Against AI Leaders Signals Escalating Backlash : A 20-year-old threw a Molotov cocktail at OpenAI CEO Sam Altman's home after writing about extinction fears from AI development, followed by a second apparent attack days later. An Indianapolis councilman also reported shots fired at his home with a "No Data Centers" note after supporting AI infrastructure. This marks a dangerous shift from peaceful AI resistance to physical attacks targeting both executives and local officials. Watch whether AI companies increase security measures and how this affects the tenor of public AI safety debates, particularly as some critics push back against framing the broader AI safety movement as extremist.

Snap Cuts 16% of Workforce in Profitability Push : Snap CEO Evan Spiegel announced layoffs of roughly 1,000 employees, or 16% of global headcount, while also closing 300 open roles. The memo cited both cost reduction goals and AI-driven productivity gains that let remaining employees "move more quickly." This is the starkest admission yet from a social platform that AI tools are directly replacing human workers at scale. Watch whether other mid-tier tech companies follow with similar cuts justified by AI productivity, and whether Snap can actually reach profitability or if this is a precursor to strategic shifts.

Meta Commits 1 Gigawatt to Broadcom Custom Chips : Meta and Broadcom announced a multiyear deal extending through 2029 for Meta to deploy 1 gigawatt of custom MTIA AI accelerators initially, scaling to multiple gigawatts by 2027. The chips will use a 2 nanometer process and are designed to reduce Meta's dependence on constrained, expensive Nvidia GPUs. This is the clearest signal yet that hyperscalers are serious about alternatives to general-purpose AI chips, even as Meta continues buying millions of Nvidia units. Watch whether Meta's internal-only ASIC strategy delivers cost advantages that Google and Amazon can match with their customer-facing chip programs.

UK Tests Show AI Model Completes Multi-Step Cyberattack : The UK's AI Security Institute found that Anthropic's Mythos Preview became the first model to complete its 32-step simulated corporate network infiltration from start to finish, succeeding in 3 of 10 attempts with an average of 22 steps completed. Previous models maxed out around 16 steps. The evaluation suggests Mythos can autonomously attack "small, weakly defended" systems but notes the tests lack active defenders and real-world detection mechanisms. Watch whether well-defended critical infrastructure proves resilient or if future models matching Mythos create an offensive-defensive AI arms race in cybersecurity.

Science Corp Prepares First Human Brain Sensor Trial : Max Hodak's Science Corporation enlisted Yale neurosurgeon Murat Günel to lead trials for a biohybrid brain-computer interface combining lab-grown neurons with electronics, with plans to surgically implant the first sensor in a human brain. The device sits on top of the brain rather than penetrating tissue and uses 520 electrodes in a pea-sized package. Unlike Neuralink's direct tissue insertion, this approach avoids brain damage from metal probes but is initially vision-only without touch or force sensing due to lack of training data. Watch whether the company can progress from basic sensors to full neuron-embedded interfaces and whether inspection applications with the $1.5 billion company's existing vision product PRIMA provide the data needed for more complex manipulation tasks.

UK Warned Big Tech Dependency Is Security Risk : The Open Rights Group published a report arguing that Britain's reliance on US tech giants across critical infrastructure creates national security exposure through laws like the CLOUD Act that enable foreign data access and service shutdowns. The group cited Microsoft allegedly cutting ICC-related email and banking services during US sanctions and estimated the UK overspends at least £500 million annually on cloud services due to vendor lock-in. Politicians across parties backed the findings, with calls for digital sovereignty through more open source software and domestic capability. Watch whether the UK government acts on the recommendations or continues awarding contracts to US providers like Palantir, and whether other European nations adopt similar sovereignty frameworks.

Scanning the Wire

Oracle commits to 2.8 GW of fuel cells from Bloom Energy : With grid connections slow and turbines scarce, Oracle is buying on-site power generation to keep datacenter expansion moving forward. (The Register)

UK fusion research gets £2.5 billion and a 2030 roadmap : Britain's atomic energy authority published technical targets for commercial fusion development before the decade ends. (The Register)

Google DeepMind ships Gemini Robotics-ER 1.6 reasoning model : The new version shows improved spatial and physical reasoning over its predecessor, advancing AI's ability to understand and manipulate the physical world. (Google DeepMind)

Spotify enters physical book sales in US and UK : The streaming platform partnered with Bookshop.org to sell print books while expanding its Page Match tool to 30 additional languages. (TechCrunch)

US late-stage venture funds raised record $23.6 billion year-to-date : The total exceeds any full year in the past 12 years as AI boom drives investor appetite for growth deals. (Wall Street Journal)

Asia startup funding jumped 93% to $27.4 billion in Q1 : Chinese startups captured $16.5 billion of the total, marking the highest quarterly investment in three years. (Crunchbase News)

ASML beats estimates and raises 2026 sales forecast : The chipmaking equipment maker reported €8.8 billion in Q1 sales and lifted full-year guidance to €36-40 billion from €34-39 billion. (CNBC)

Uber racing to spend over $10 billion on robotaxis : The ride-hailing company is committing $7.5 billion to vehicle purchases and $2.5 billion to equity stakes in autonomous vehicle makers. (Financial Times)

YouTube will pause livestream ads during peak engagement : The platform is holding back ad breaks when viewer interaction spikes to protect creator momentum. (TechCrunch)

IBM settles for $17 million under Trump diversity initiative : The company became the first to pay up under the administration's Civil Rights Fraud Initiative without admitting liability. (The Register)

Outlier

Google's Personalized AI Reaches India: Google brought Gemini Personal Intelligence to India, letting users connect Gmail and Photos for personalized answers. This marks the first major deployment of account-integrated AI outside Western markets, testing whether personalized AI can scale in a country where data privacy norms, digital literacy levels, and infrastructure constraints differ dramatically from the US and Europe. If Indian users embrace having AI parse their email and photos for convenience, it suggests the privacy-for-utility tradeoff that Western tech critics obsess over may not be universal. If they don't, it reveals that personalized AI's value proposition depends on assumptions about digital trust that don't export cleanly. Either outcome tells us something important about whether AI's next billion users will adopt or reject the surveillance-as-service model.

The mirror bacteria researchers changed their minds when they realized they were building something their own immune systems couldn't recognize. Most of us are building things we can't fully understand either, just at smaller scales. Sleep well.