The Infrastructure Play
The Infrastructure Play
The real battle in AI isn’t over models anymore. It’s over who controls the stack that runs them. This week revealed the contours of a fundamental shift: the companies winning aren’t just building better LLMs, they’re building the infrastructure, standards, and integrations that make AI useful at scale. From TikTok’s forced restructuring to OpenAI’s ecosystem bets to Anthropic’s open standards play, the pattern is unmistakable. Whoever owns the plumbing wins the game.
Deep Dive
TikTok’s Sale Closes the China Question but Opens a Bigger One
The deal is finally happening. TikTok US will close on January 22 as a joint venture with Oracle handling data security, Silver Lake and MGX each taking 15%, and ByteDance retaining 19.9% while 30.1% stays with existing investors. On the surface, this solves the divest-or-ban crisis that nearly killed the platform. Underneath, it reveals something more important: the U.S. has accepted that national security can be managed through infrastructure control rather than ownership.
The real story isn’t that ByteDance lost TikTok. It’s that Oracle’s role fundamentally rewrites how American tech policy works. By insisting on Oracle as a “trusted security partner” overseeing algorithm retraining, data storage, and content moderation, the U.S. government essentially turned a data company into a regulatory apparatus. Oracle gets a license to inspect TikTok’s recommendation engine in real time. This is infrastructure as policy. The algorithm can’t be “outside manipulated.” American user data can’t leave American clouds. The U.S. didn’t seize the company, it seized the stack.
This precedent will matter for every other platform facing national security scrutiny. It’s a template that sidesteps the complexity of forced sales or full bans. You keep the business, you lose the autonomy. The question for China and TikTok’s parent: Is 19.9% of a functioning platform better than 0% of a banned one? For now, yes. But the window for renegotiating that deal after January 22 will be zero.
OpenAI’s $100B Fundraise Is Really a War for Compute Dominance
OpenAI wants \(100 billion at an \)830 billion valuation. That’s not about proving the model works anymore. It’s not even really about funding research. It’s about guaranteeing access to the infrastructure that will be scarce for the next five years: GPUs, data center capacity, and power. The company that can’t reliably source compute dies. The company that can controls pricing, latency, and feature velocity.
The fundraising round reflects this reality. OpenAI needs capital not because its business model is uncertain but because scaling LLM inference is capital-intensive in ways that don’t show up in traditional venture returns. Each percentage point improvement in inference efficiency matters more than each percentage point improvement in model quality. The compute bill is the budget. This is why OpenAI keeps talking about reasoning and why it’s willing to accept a 120-day training window for o1 responses. The real constraint is power, not talent.
Watch who joins this round. If it’s primarily sovereign wealth funds and energy companies rather than traditional VCs, that confirms the thesis: this is an infrastructure play being funded like infrastructure. The companies that own or control power generation and data center capacity will win. Sam Altman knows this. His push to partner with nuclear fusion startups and his recent statements about energy aren’t tangential to AI strategy. They’re core to it.
Anthropic’s Skills Strategy: Beating OpenAI by Not Fighting It
Anthropic released Agent Skills as an open standard, not a proprietary feature. This looks like generosity. It’s actually strategic dominance.
By opening the specification, Anthropic handed OpenAI and Google a blueprint. OpenAI has already replicated it. But here’s what happened: the company that gave away the technology now owns the ecosystem. Fortune 500 companies are building skills for Claude. 20,000 community-created skills exist on GitHub. Microsoft integrated Agent Skills into VS Code. When OpenAI adopted structurally identical architecture, they didn’t just copy a feature. They validated Anthropic’s bet that procedural knowledge encoded as modular, portable toolkits is the future of AI work.
The real win isn’t that other companies can’t use skills. It’s that by open-sourcing the standard, Anthropic made it impossible for OpenAI or Google to lock companies into proprietary implementations. The skills economy now runs on Anthropic’s spec. ChatGPT uses skills that can be moved to Claude. Claude users can adopt skills built for other platforms. This is a glimpse at how open standards function as competitive weapons in network markets. The player who sets the standard doesn’t need to own the whole network to extract value from it.
The second-order effect: companies spending engineering time building and curating skills are making switching costs higher, not lower. A Fortune 500 company with 300 custom-built skills for Claude faces real friction moving to ChatGPT, even if ChatGPT is slightly better at some tasks. Anthropic didn’t lock in customers through proprietary models. It locked them in through procedural knowledge that lives in their organization.
Signal Shots
Trump Media Acquires Fusion Power Company in $6B+ Merger — Trump Media and Technology Group announced a merger with TAE Technologies, a fusion energy startup, in a deal valuing the combined entity at $6 billion or more. The merger positions Trump Media as a diversified holding company rather than a pure-play social media company. This is either a visionary play on the coming energy crisis that AI will trigger or a billionaire’s vanity project. Watch whether Trump Media actually contributes engineering or capital to fusion development or if this is purely a financial engineering move designed to prop up TMTG’s valuation.
Hut 8 and Fluidstack Building AI Data Center for Anthropic in Louisiana — Mining infrastructure company Hut 8 and GPU cloud provider Fluidstack are building a data center in Louisiana specifically designed for Anthropic’s infrastructure needs. This signals that the tight GPU market is pushing even frontier AI companies toward bespoke infrastructure rather than relying on hyperscaler capacity. For Anthropic, it reduces dependency on AWS compute pricing and volatility. For Hut 8, it’s a long-term revenue stream and a signal that crypto mining’s infrastructure becomes AI infrastructure.
House Passes SPEED Act for AI Infrastructure Permitting — The House passed the SPEED Act, which eases permitting processes for AI data centers and infrastructure projects. This moves to the Senate, where permitting reform is already in early discussions. The bottleneck for AI deployment isn’t chips or talent. It’s zoning, environmental review, and grid connection. This is the non-glamorous but essential piece of infrastructure policy. Watch whether the Senate passes a meaningful version or if it gets watered down by local opposition.
China Launches CENI: Its Answer to ARPANET — Beijing certified the China Environment for Network Innovation, a vast experimental research network intended to cement China’s position in networking research. The framing as “heir to ARPANET” is telling. China isn’t trying to beat the U.S. on individual technologies anymore. It’s trying to build a parallel stack. Watch whether CENI becomes a genuine alternative substrate for research that can’t happen on Western networks or if it remains largely symbolic.
U.S. Restricts Investment in Chinese Tech via NDAA — The National Defense Authorization Act that Trump signed includes provisions cutting capital flows to Chinese military and surveillance tech companies. This completes the bifurcation of the global tech stack. American capital can no longer fund certain categories of Chinese tech development. Chinese capital restrictions on Western tech are becoming reciprocal. The world is splitting into incompatible infrastructure regimes.
Nvidia H200 Export Approval Based on Flawed Premise — The Trump administration approved exports of Nvidia’s H200 chips to China based on the belief that Huawei is a viable competitor. Council on Foreign Relations analysis shows the gap between Nvidia and Huawei is actually widening, not narrowing. This suggests the export decision was political theater, not strategic calculation. If Huawei isn’t catching up, H200 exports accelerate China’s deficit rather than level competition.
Scanning the Wire
OpenAI Launches ChatGPT App Store — OpenAI opened submission windows for third-party developers to build ChatGPT-integrated applications, echoing Apple’s app store model but without the revenue split chaos. This is ecosystem lock-in disguised as openness.
ChatGPT Mobile App Hits $3B in Consumer Spending — ChatGPT’s mobile app crossed $3 billion in lifetime consumer spending in 31 months, faster than TikTok and major streaming apps. Consumer willingness to pay for AI is proven. The question is whether those margins persist once competition intensifies.
Sam Altman on OpenAI’s “Code Red” Call — In an interview with Big Technology, Sam Altman discussed OpenAI’s enterprise strategy, IPO plans, and product ambitions. The “code red” framing around competitive threats signals internal conviction that the lead is narrowing faster than public statements suggest.
Amazon’s Alexa+ Adds Video Understanding to Ring Doorbells — Amazon deployed conversational AI to Ring that can identify people based on uniforms and actions. This is the integration play: existing hardware getting smarter without replacement cycles.
Optus Firewall Upgrade Failures Caused Two Deaths — An Australian telco’s botched firewall upgrade knocked out emergency services routing and contributed to two deaths. Ten mistakes in execution, poor escalation protocols, and bad documentation. This is why infrastructure reliability matters more than feature velocity.
Snowflake Global Update Caused Cascading Failures — A Snowflake software update caused operational failures across 10 of 23 regions. When data infrastructure breaks, everything downstream breaks. This is the fragility hiding under the AI boom.
YouTube Bans Channels Making Fake AI Movie Trailers — YouTube removed two popular channels that created AI-generated movie trailers, despite Google’s enthusiasm for AI content elsewhere. The policy inconsistency reveals that Google’s AI strategy is still being written in real time.
LLMs Boost Scientific Publication Volume but Not Quality — A study found that when researchers use LLMs, paper output increases but quality metrics stagnate. More content, less rigor. This pattern will repeat across every field that adopts LLM-assisted workflows.
School Security AI Mistakes Clarinet for Gun — A Florida school’s AI security system flagged a clarinet as a firearm, triggering lockdown. Human review didn’t catch the error. Schools are expanding AI deployment despite demonstrated unreliability in the one domain where false positives have life-or-death consequences.
Instacart Pays $60M to Settle FTC Consumer Deception Claims — Instacart agreed to settle FTC allegations of misleading fee disclosures and denied refunds. This is the emerging pattern: platform companies get caught, pay settlements, move on. The incentive structure isn’t changing behavior.
Meta Developing Mango AI Model for Images and Video — Meta is building a multimodal model codenamed Mango expected in H1 2026, alongside a new LLM called Avocado. Meta is still trying to compete in the model layer despite having fewer compute resources than OpenAI or Google. Watch whether Llama 4’s performance justifies the capital spend.
Outlier
FCC Commissioner Brendan Carr Says Agency “Isn’t Independent” — A Republican FCC commissioner testified before Congress that the FCC “isn’t independent,” raising alarm about how the Trump administration might weaponize the agency. This signals that regulatory capture isn’t a future risk anymore; it’s the current state. When the people supposed to check executive power acknowledge they can’t, the infrastructure that constrains big tech disappears. We’re entering a period where platform power grows unchecked by formal regulatory constraints.
See you on Monday when the new year quietly pivots everything you thought you knew about AI timelines, energy constraints, and whether any of this actually matters if the grid can’t handle it.