Dear Dysruptors,
Fernando Santa Cruz here in the 45° edition of Synapsis Weekly, where Claude surpassed OpenAI in corporate spending for the first time, Google and SpaceX began negotiating orbital servers because the terrestrial power grid can no longer keep up, and Anthropic launched in the same week a classified model that finds vulnerabilities in macOS and a package of connectors so a small shop owner can automate her invoicing.
Last week, the map was redrawn.
The AI frontier moved to the sky.
Practical AI came down to the corner store.
Anthropic stopped being OpenAI’s challenger and became the world’s leading corporate AI provider. Google and SpaceX started talking seriously about orbital servers because 49,000 homes in Lake Tahoe are already giving up 75% of their electricity to data centres.
And at the same time, Anthropic launched ready-made connectors for QuickBooks, Salesforce, Shopify, and HubSpot.
The same lab that has a government-classified model also launched last week the version for the owner of a small business.
This newsletter digs deeper into the WhatsApp summaries (week of May 11 to 16) to understand why Anthropic won B2B by being “boring,” why the AI frontier stopped being about code and became about thermodynamics, and why the most practical news for an SMB last week came from none of the frontier headlines.
Anthropic Captures 52% of Corporate Spending and OpenAI Falls to 41%: The Age of Boring AI Has Begun
The Ramp AI Index for May confirmed something that changes the conversation.
Claude captures 52% of corporate spending. OpenAI fell to 41%.
It’s the first time a model other than ChatGPT has led the B2B market.
The catalyst was the PwC alliance to certify and deploy Anthropic Claude across 360,000 employees, with legal connectors that reduce contract review times by up to 80%.
Mega-corporations didn’t buy the smartest model.
They bought the most predictable one.
Zero hallucinations in critical scenarios. Zero use of client data for training. Corporate auditability.
It’s like when a bakery hires a new employee. They’re not looking for the most creative one. They want someone who shows up on time, follows the recipe, and doesn’t invent ingredients.
Intelligence is marketing.
Predictability is revenue.
For an SMB, this reorders the question. It’s no longer “which model is smartest?” It’s “which one makes fewer mistakes in my critical decisions?”
Question for your strategy: In which critical business processes are you choosing the “smartest” model when what you actually need is the most boringly reliable one?
49,000 Homes Give Up 75% of Their Electricity and Google Starts Negotiating Orbital Servers
NV Energy redirected 75% of the power supply from 49,000 residents in Lake Tahoe to data centres in Nevada. Residential rates jumped 77%.
Entire households giving up their power so AI models can reason.
In response, Google and SpaceX began serious talks to deploy AI servers in low Earth orbit. Starlink laser links exceeding 100 Gbps. Free cooling in the vacuum of space. Solar power around the clock.
Let’s sit with that for a moment.
The bottleneck of AI stopped being code. It stopped being chips.
It became thermodynamic physics.
It’s like when a colonial city grew until the drainage system collapsed. They didn’t build a new city. They redesigned the water infrastructure underground.
Except this time, the redesign is above our heads.
For an SMB, this isn’t a distant concern. It means growing geographic volatility in inference pricing. And it means increasing weight on tools that run small models directly on the device itself, like WhatsApp with Muse Spark.
Energy stopped being an invisible cost.
It’s a supply chain risk.
Question for your operations: How many of your critical tools depend on cloud inference, and what happens to your business if prices rise 30% due to energy constraints next year?
Recursive Raises $650 Million to Build AI That Rewrites Itself: $4.65 Billion Valuation for Zero Product
The startup Recursive Superintelligence announced a $650 million round. Valuation of $4.65 billion.
Zero commercial product. Zero customers.
A single bet. Build AI capable of autonomously rewriting superior versions of its own training code. Without human supervision.
If it works, the next leap in AI won’t come from an app. Not from a bigger model.
It will come from a model that programmes other models.
It’s like when a chef hires a sous-chef who improves the recipe without asking permission, every night, until the restaurant owner no longer knows what’s being cooked in their own kitchen.
The philosophical detail is unsettling.
Silicon Valley is betting $650 million that data generated by humans is no longer sufficient to evolve AI.
We as a species are being, literally, removed from the loop.
For SMBs and entrepreneurs, this isn’t a technophobic reading. It’s a strategic one. Over the next five years we could see AI capabilities that no human researcher designed.
This changes how you evaluate a vendor. Not just “what does it do today?” but “who decides what it will do tomorrow?”
Question for your three-year plan: Are you choosing AI vendors based on their current capabilities, or on their transparency about how and who decides the future evolution of the model?
Gemini 3.2 Flash Reaches 92% of GPT-5.5 at 20% of the Cost: The Era of Expensive Tokens Is Almost Over
Industry leaks confirmed via X place Gemini 3.2 Flash at 92% of GPT-5.5’s reasoning performance. Cost 20 times lower. Latency under 200 milliseconds.
If the leak is confirmed, the main economic barrier to deploying agents at scale falls.
What costs $1,000 a month in inference today could cost $50 a month from now.
For two years, SMBs were locked out of autonomous agents because each action cost too much to justify the ROI.
If one query cost 10 cents and an agent ran 1,000 a day, that was $100 daily in tokens alone.
With Flash, that same workday would cost $5.
It’s like when the internet moved from per-minute billing to flat rate. Video calls, streaming, and social media all became viable overnight.
The business model of AI’s future will be logic as a basic utility.
Like water. Like electricity.
For an SMB, this shifts the adoption question. It’s no longer “can we afford AI?” It’s “what were we doing by hand that we should now automate before the competition does?”
Question for your calendar: What three repetitive processes in your business did you dismiss last year as too expensive in tokens, and which ones come back to the table if the cost drops 95%?
Mythos Finds Zero-Days in macOS Before Any Human and OpenAI Responds with Military-Grade Daybreak
Anthropic’s classified model, Mythos, identified multiple Zero-Day vulnerabilities in Apple’s macOS.
Autonomously. Before any human researcher.
Meanwhile, OpenAI responded by launching Daybreak, a military-grade cybersecurity platform designed exclusively for governments and critical infrastructure.
The data point few are connecting.
Google reports a 300% increase in cybercriminals using LLMs to generate polymorphic malware. Malicious software that changes form to evade traditional defences.
The nature of defence has changed.
It’s no longer about scanning for known virus signatures. AI builds reasoning trees to understand the entire architecture of the software it’s attacking or defending.
It’s like moving from a security guard who checks a banned-persons list to a detective who studies the building’s blueprints to anticipate every possible entry point.
For an SMB, traditional tools (antivirus, basic firewalls) are increasingly insufficient against AI-generated attacks. Serious vendors will start including defensive AI in their packages. And prices are going to rise.
Companies that don’t update their cyber defences in the next twelve months will be easy targets.
From attacks that cost pennies to produce.
Question for your team: Does your company know what kind of AI, if any, your current cybersecurity provider is using, and what happens if attackers bring better models than your defence?
Figure AI Live-Streams Its Humanoid Robot Working 34 Hours Without an Operator and Sorting 43,000 Packages
Figure Robotics streamed live its humanoid robot working 100% autonomously in a real industrial plant.
34 continuous hours. 43,000 packages sorted.
Zero teleoperation. Zero human supervision.
This is an operational milestone, not just a technical one.
For years, humanoid robot demonstrations came with asterisks. Short duration. Controlled environment. Remote operator off-camera.
Not this time.
What changed wasn’t the actuators. The hardware has been ready for years. What matured were the visuomotor neural networks. AI learned to coordinate vision and physical movement without static programming.
It’s like when a new server goes from needing instructions for every table to reading the entire room’s dynamic and deciding on their own.
For an SMB in logistics, distribution, manufacturing, or care, this isn’t relevant in six months. But the commercial adoption curve is starting its countdown.
The advantage won’t be having robots.
It will be having processes clear enough for a robot to execute them.
Question for your sector: What physical processes in your business are so clearly documented today that you could hand them off to someone, human or otherwise, without corridor explanations?
Thinking Machines Lab s Launches AI That Interrupts You When You’re Wrong: Human Imperfection as a Design Standard
Mira Murati’s startup (former OpenAI CTO) presented interaction models that break the turn-based chat format.
Real-time simultaneous translation. The AI interrupts when it detects an error. It talks over the user when necessary.
They deliberately adopted the “flaws” of human conversation.
Pauses. Overlaps. Abrupt cognitive redirections.
The opposite of the perfect, polite AI the industry has chased for years.
And the result is paradoxical.
It generates more psychological adoption than any robotically perfect response.
It’s like when a mentor you trust cuts you off mid-sentence because they’re about to say something important. You’re not annoyed. You trust them more, not less.
Deep AI adoption won’t come from machines that imitate the “best” of humans.
It will come from machines that imitate the “authentic.” Including the imperfect.
For an SMB thinking about building a customer service agent, this changes the brief. The goal is no longer “make it seem like a polite human.” It’s “make it behave like a useful human.”
A useful human interrupts. A useful human corrects.
Question for your product team: Is the AI your company uses or develops designed to please the user, or to correct them when it’s useful to do so, even if it’s less comfortable?
Claude Arrives at the Corner Store: Ready-Made Connectors with QuickBooks, Salesforce, Shopify, and HubSpot
Anthropic launched Claude for Small Business.
The same lab that has the Mythos model classified for national security risk.
And the package is direct. Pre-built connectors with QuickBooks. Salesforce. Shopify. HubSpot. Google Workspace.
No code. No engineers. No APIs.
An SMB can automate invoicing, data entry, customer service, quote generation, and monthly reports.
It’s the most direct bet by a frontier lab to make Claude a tool for small businesses, not just large ones.
It’s like when banks moved from serving only large corporations to opening branches for micro-entrepreneurs. Not out of philanthropy. Because they discovered that loose volume is worth more than large accounts poorly served.
But the strategic detail is subtler.
The ready-made connectors eliminate the most expensive adoption barrier in AI. Not the price of the model.
The technical consulting required to connect it to your systems.
For an SMB in Mérida, Monterrey, Toronto, or Buenos Aires, this means a week of testing instead of a three-month project.
The frontier is no longer in San Francisco.
It’s inside QuickBooks. The only thing left is to activate it.
Question for your adoption: What are the three tools your business uses every day, and which of them already has a ready-made connector with Claude, ChatGPT, or Gemini that nobody in your company has tried?
Tools You Can Use on Monday
- Claude for Small Business – Anthropic’s package with pre-built connectors for QuickBooks, Salesforce, Shopify, HubSpot, and Google Workspace. Automates invoicing, customer service, and data entry without code. The most direct path to the frontier without hiring engineers.
- WhatsApp Incognito Chat with Meta AI – Meta activated on-device processing with the Muse Spark model (2 billion parameters). Queries don’t leave the phone. Corporate-grade privacy in the app everyone already uses.
- Google AI Pointer by DeepMind – Intelligent cursor that replaces the traditional mouse. Understands the screen visually, suggests proactive actions, and automates multi-step sequences by highlighting elements. Reduces click fatigue by up to 60%.
- TikTok MCP for Advertising Agents – MCP server that lets you delegate complete campaign management to AI agents. Design, budgets, and real-time bidding without human oversight. Brands report 35% higher ROI.
- Krea 2 for Brand Visual Consistency – Foundation model with parametric control over styles, lighting, and character consistency. Maintains the same look across an entire campaign, not across isolated images.
My Invitation This Week: Monday’s Blank Page
One sheet. Three columns. Twenty minutes.
That’s the whole exercise.
Take a piece of paper. Draw three vertical columns. Write these headings at the top: Tool, Colour, Process.
- In the first column, list everything the team opens every day without thinking. CRM, accounting, e-commerce, calendar, email, design, inventory, ad platform. No filtering or ranking. Just write.
- In the second column, next to each tool, assign a colour. Green if it has a native connector with Claude, ChatGPT, or Gemini (QuickBooks, Salesforce, Shopify, HubSpot, Notion, Slack, Microsoft 365, and Google Workspace are already green by default). Yellow if there’s an integration via Zapier or Make. Red if it requires custom development.
- In the third column, only for the green ones, write in five words what repetitive task Claude could handle last week. “Generate invoices from orders.” “Summarise CRM conversations.” “Sort emails by urgency.” “Sync inventory with Shopify.”
That’s it.
We’re not installing anything. Not hiring anyone. Not designing a project.
We’re just lifting our eyes to see the map that already exists.
The uncomfortable discovery is almost always the same. Between 40% and 60% of the team’s tools show up green. The frontier entered the business months ago. Nobody had just pulled out the sheet to see it.
And that sheet, folded and stuck to the monitor, is worth more last week than any of the tools listed on it.
Closing
This wasn’t a one-announcement week.
It was the week of the great reordering.
The AI frontier moved to space. Practical AI came down to the corner store.
The lab that has a model classified for national security risk also launched the version for a taco shop owner to automate her invoicing.
SMBs and entrepreneurs building in this era have a short and rare window. The frontier is accessible. The connectors are ready. And the large competition hasn’t moved yet.
Because the question that defines who will build resilient businesses isn’t whether AI will arrive.
It’s whether we will activate what’s already inside the tools we already use.
Fernando Santa Cruz Head of AI & Automation @ Adivor Consulting
