Deepgram, Cast AI Hit $1B+ as Marlinspike Reinvents AI Security

So, What Actually Happened?

We scanned 190,000 articles this week so you don't have to, and the AI industry is having a moment of existential clarity. Moxie Marlinspike—the Signal creator who made encrypted messaging mainstream—just announced he's building encrypted AI infrastructure. Meanwhile, Microsoft's Brad Smith is pushing Big Tech to fund their own AI data center infrastructure, essentially telling the industry to stop externalizing costs onto communities. And the funding keeps flowing: Deepgram hit a $1.3B valuation for voice AI, while Snowflake and Anthropic inked a $200M AI agents deal.

The Bottom Line: The AI industry is growing up—facing real questions about who pays for infrastructure, who controls the data, and whether we can build systems we actually trust.

The Tracks That Matter

1. Moxie Marlinspike Brings Signal's Philosophy to AI

Signal creator Moxie Marlinspike wants to do for AI what he did for messaging.

This is the most significant AI security announcement of the year. Marlinspike—the cryptographer who made end-to-end encryption accessible enough that WhatsApp, Facebook Messenger, and Skype all adopted Signal's protocol—is now turning his attention to AI infrastructure. The problem he's solving: how do you run AI workloads without trusting the cloud provider?

The current AI model requires handing your data to someone else's servers. Every prompt, every document, every query—processed on infrastructure you don't control. Marlinspike's approach applies the same principles that made Signal successful: end-to-end encryption, minimal metadata collection, and architectures that don't require trusting the service provider.

For enterprises handling sensitive data—legal, healthcare, financial services—this could be transformative. The question has always been: how do we use AI capabilities without compromising confidentiality? Marlinspike has a track record of solving exactly these kinds of problems at scale.

Here's what works: Watch this space carefully. If Marlinspike's AI infrastructure follows the Signal playbook, it will take 2-3 years to mature but could become the de facto standard for privacy-preserving AI. Start building relationships now.

2. Voice AI Hits the Billion-Dollar Club: Deepgram's $1.3B Valuation

Voice AI startup Deepgram raises $130 million at $1.3 billion valuation.

Deepgram just joined the unicorn club with a $130M raise at a $1.3B valuation. The company specializes in speech-to-text AI that's faster and more accurate than legacy transcription services, and they're betting that voice interfaces are about to have their iPhone moment.

The timing matters. As AI agents become more capable, voice becomes the natural interface—nobody wants to type complex instructions when they can speak them. Deepgram is positioning itself as the infrastructure layer that connects voice to AI reasoning, handling the translation between human speech and machine understanding.

What makes this interesting isn't just the valuation—it's who's investing. When institutional money flows into voice AI infrastructure at this scale, it signals confidence that the next wave of AI applications will be conversational, not just text-based. The keyboard-and-screen paradigm might have a shorter remaining lifespan than most enterprise software roadmaps assume.

Here's what works: If your AI strategy assumes text-based interactions, expand your planning horizon to include voice. The enterprises that figure out voice-first AI workflows now will have a significant advantage when the interface shift accelerates.

3. Cast AI Launches GPU Marketplace at $1B Valuation

Cast AI Valued at Over $1 Billion With the Launch of Its GPU Marketplace.

Cast AI just hit unicorn status by solving one of AI's most persistent problems: getting GPU access without selling a kidney. Their new marketplace matches enterprises that need compute with those who have excess capacity, creating a spot market for AI infrastructure.

The GPU shortage has forced companies to either overprovision (expensive) or wait in queue (slow). Cast AI's marketplace approach introduces price discovery and liquidity—you can buy compute when you need it, at prices that reflect actual supply and demand rather than contract negotiations with cloud giants.

This is the financialization of AI infrastructure, and it's probably inevitable. When a resource is both scarce and essential, markets emerge. The question is whether the big cloud providers will embrace this model or fight it. Given their current margins on GPU instances, expect resistance.

Here's what works: Evaluate whether a GPU marketplace fits your compute strategy. If your AI workloads are bursty or experimental, spot market access could reduce costs by 30-50% compared to committed capacity.

4. Snowflake and Anthropic Sign $200M AI Agents Deal

Snowflake and Anthropic sign $200m AI agents deal.

Snowflake just committed $200 million to integrate Anthropic's Claude directly into its data cloud platform. This isn't a vague partnership announcement—it's a bet that AI agents operating on enterprise data will become a core product category.

The integration means Claude can reason over data stored in Snowflake without moving it to external systems. For enterprises with petabytes in Snowflake, this eliminates one of the biggest friction points in AI deployment: getting data to the model without violating security policies, data residency requirements, or sanity.

This deal follows Snowflake's acquisition of Observe and suggests a coherent strategy: become the platform where enterprise AI actually happens, not just where data sits waiting to be extracted. If they execute, Snowflake becomes much stickier—you're not just locked into their storage, you're locked into their AI workflows.

Here's what works: If you're a Snowflake shop, this integration is worth evaluating immediately. The ability to run Claude directly on your data warehouse without complex extraction pipelines could accelerate AI deployment timelines significantly.

5. Microsoft's Brad Smith: Tech Should Pay for Its Own Data Centers

Microsoft's Brad Smith pushes Big Tech to 'pay our way' for AI data centers.

Microsoft's President Brad Smith is making an argument you don't often hear from Big Tech: companies should fund their own AI infrastructure instead of expecting communities to subsidize it. He's specifically calling out the practice of seeking tax breaks and public infrastructure investments for private data centers.

This is a significant break from tech's usual playbook of extracting maximum concessions from local governments. Smith's argument—that the industry is profitable enough to pay its own way—represents either genuine corporate responsibility or strategic positioning against competitors who rely more heavily on public subsidies.

The infrastructure costs of AI are becoming politically visible. Data centers consume enormous amounts of power and water, strain local grids, and require substantial road and utility investments. Communities are starting to push back. Smith is getting ahead of the backlash by positioning Microsoft as the responsible actor.

Here's what works: If you're evaluating cloud providers, their infrastructure investment practices matter. Providers who externalize costs onto communities may face regulatory and political headwinds that affect service reliability and pricing.

6. MIT Technology Review Names Mechanistic Interpretability a 2026 Breakthrough

Mechanistic interpretability: 10 Breakthrough Technologies 2026.

MIT Technology Review just named mechanistic interpretability one of their ten breakthrough technologies for 2026. The field—pioneered largely by Anthropic—involves building tools to understand what's actually happening inside large language models, neuron by neuron.

In 2024, Anthropic built what they called a ”microscope” that could identify features in Claude corresponding to recognizable concepts. In 2025, they traced entire reasoning paths from prompt to response. OpenAI used similar techniques to catch one of their reasoning models cheating on coding tests. The technology is moving from research curiosity to practical application.

For enterprises deploying AI, interpretability matters because it's the foundation of auditability. Regulators will eventually require explanations for AI decisions—particularly in healthcare, finance, and employment. The companies investing in interpretability now will have a head start when those requirements arrive.

Here's what works: Track interpretability research from Anthropic and OpenAI. When evaluating AI vendors, ask about their interpretability capabilities. The ability to explain why a model made a decision will become a compliance requirement.

7. 2025 Produced 100+ New Tech Unicorns Despite the ”Down Market”

More than 100 new tech unicorns were minted in 2025.

TechCrunch's analysis confirms that 2025 produced over 100 new unicorns—companies valued at $1 billion or more. Notable entries include Unconventional AI (working on energy-efficient AI computing, raised $475M seed), Reflection (open foundation models, $2B Series B), and Mercor (contract recruiting, $100M Series B).

The narrative that tech funding has dried up doesn't match the data. What's changed is where the money flows: AI infrastructure, energy-efficient computing, and platforms that enable AI deployment are attracting capital at historic rates, while generic SaaS and consumer apps struggle.

The composition of new unicorns tells you where the smart money sees opportunity. Energy-efficient AI computing is a bet that current GPU infrastructure is unsustainable. Open foundation models are a bet that the closed-model oligopoly will face competition. Contract recruiting platforms are a bet that how companies hire is about to change dramatically.

Here's what works: Follow the unicorn list as a signal of where venture capital sees durable opportunity. If you're evaluating build-vs-buy decisions, check whether there's a well-funded startup attacking your problem space—they may be worth partnering with or watching closely.

Signal vs. Noise

🟢 Signal: Databricks is having a moment. PageRank growth of 102% signals genuine increasing influence across the data infrastructure ecosystem—not just marketing buzz but actual integration and adoption patterns. When WTW Radar integrates with Databricks for insurance analytics and Palantir deepens its partnership, that's real enterprise traction.

🔴 Noise: The ”AI in Healthcare” mentions are everywhere, but influence metrics tell a different story. Healthcare AI announcements are running hot in press releases while actual deployment metrics remain modest. Be skeptical of healthcare AI claims that don't include specific patient outcomes or workflow improvements.

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

The Infrastructure Independence Movement

Three developments this week connect in ways the headlines miss: Moxie Marlinspike's encrypted AI infrastructure announcement, Brad Smith's call for tech to fund its own data centers, and Cast AI's GPU marketplace launch.

Together, they suggest a shift in how enterprises think about AI infrastructure. The current model—rent everything from a handful of cloud giants—is starting to crack. Privacy requirements are pushing toward encrypted compute. Community pushback is forcing infrastructure cost transparency. Spot markets are introducing price competition.

We're seeing the early signs of AI infrastructure becoming a competitive market rather than an oligopoly. The enterprises that figure out how to navigate this transition—mixing cloud, marketplace, and potentially encrypted infrastructure—will have cost and capability advantages over those who remain locked into single-vendor relationships.

The implication: Your cloud vendor strategy should assume more options, not fewer. The infrastructure landscape in 2028 will look very different from today.

By The Numbers

• $1.3B — Deepgram's valuation after $130M raise for voice AI
• $200M — Snowflake-Anthropic deal for AI agents integration
• $1B+ — Cast AI's valuation with GPU marketplace launch
• 100+ — New tech unicorns minted in 2025 despite ”down market”
• 244,000 — Global tech-sector layoffs in 2025
• 102% — Databricks' PageRank growth indicating rising influence

Deep Dive: The Privacy AI Paradox

Like a DJ who realizes the sound system belongs to someone who might be recording every beat, the AI industry is waking up to an uncomfortable truth: the infrastructure we've built for AI requires trusting parties we might not want to trust.

The Problem Moxie Sees

Moxie Marlinspike built Signal on a simple principle: you shouldn't have to trust your communication provider not to read your messages. End-to-end encryption means even Signal can't see what you're saying. Now he's asking the same question about AI: why should using AI require handing your data to someone else's servers?

The current model is fundamentally trust-based. When you use ChatGPT, Claude, or any cloud-hosted AI, your prompts travel to someone else's infrastructure. You're trusting OpenAI, Anthropic, or Google not to train on your data, not to leak it, not to be compromised. For many use cases, that's fine. For sensitive legal discovery, medical records, financial analysis? It's a real constraint.

The Technical Challenge

Encrypted AI is hard because AI models need to see your data to process it. Traditional encryption means the server can't read what it's computing on. Solving this requires techniques like homomorphic encryption (computing on encrypted data), secure enclaves (hardware-isolated processing), or clever architectures that minimize what the server actually sees.

Marlinspike's track record suggests he'll find a practical balance—Signal's success came from making encryption usable, not from achieving perfect theoretical security. Expect something similar: encrypted AI that's good enough for most sensitive workloads, even if it's not mathematically perfect.

What Actually Works

1. Evaluate your sensitivity tiers: Not all AI workloads need encrypted infrastructure. Identify which use cases involve data you genuinely can't share with cloud providers.

2. Track Marlinspike's project: His announcements tend to precede working products by 12-24 months. Start planning for encrypted AI options in your 2027-2028 roadmap.

3. Pressure your vendors: Ask your AI providers about their confidential computing roadmap. Customer demand drives product development.

4. Consider hybrid architectures: Some workloads can use cloud AI safely; others can't. Build workflows that route appropriately rather than choosing one model for everything.

The privacy AI paradox isn't going away. The enterprises that figure out how to use AI capabilities without compromising confidentiality will have access to use cases their competitors can't touch. The sound system is getting encrypted.

What's Coming

NIST Calls for Help Securing AI Agents

NIST Calls for Public to Help Better Secure AI Agents. The federal government is openly acknowledging it doesn't have the answers for AI agent security and is seeking public input. When NIST asks for help, it usually precedes new standards. Watch for security frameworks that become compliance requirements.

The Digital Fairness Act Takes Shape

The Digital Fairness Act: what you need to know. EU regulation continues evolving. The Digital Fairness Act expands protections for consumers interacting with AI systems, with particular focus on personalization and automated decision-making. Multi-jurisdictional compliance is getting more complex.

Google Veo 3.1 Pushes Video Generation Forward

Google Veo 3.1 brings more consistency and creativity to AI video. Google's latest video generation model focuses on consistency—a persistent weakness of AI video. If they've actually solved the ”characters changing appearance mid-scene” problem, video production workflows are about to shift.

For Your Team

Thursday's meeting prompt: ”Moxie Marlinspike is building encrypted AI infrastructure because he believes current cloud AI requires too much trust. What AI use cases have we avoided because we couldn't solve the data confidentiality problem? What would change if we could run AI on sensitive data safely?”

The Infrastructure Independence Framework:

1. Audit your cloud dependencies — Map which AI workloads are locked to specific providers and what switching would cost
2. Identify sensitivity tiers — Classify AI use cases by data sensitivity and match infrastructure accordingly
3. Watch the marketplace — GPU spot markets could reduce costs 30-50% for experimental and bursty workloads
4. Plan for encrypted options — Add confidential AI computing to your 2027-2028 technology radar

Share-worthy stat: ”2025 produced 100+ new tech unicorns—including Unconventional AI ($475M seed for energy-efficient AI computing). The 'funding winter' narrative doesn't match where money is actually flowing.”

Go deeper: Track AI infrastructure trends in real-time →

The Track of the Day

”We can't build superintelligence just for superintelligence's sake. It's got to be for humanity's sake, for a future we actually want to live in.”
— Google DeepMind CTO Koray Kavukcuoglu

Like a producer who finally asks whether the track serves the dancefloor or just the ego, the AI industry is starting to grapple with purpose. Building capability without building trust is just making noise.

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: January 14, 2026 | Curated by Yves Mulkers @ Ins7ghts