Four stories driving the conversation today. Intel just raised 20 billion dollars in a common stock offering to fund its AI foundry ambitions, pricing at 95 dollars per share. AMD, meanwhile, bought a Toronto startup called Taalas that etches AI models directly into silicon. Riot Platforms signed a 20-year, 9.1 billion dollar compute deal with Anthropic, leasing 191 megawatts at its Rockdale, Texas campus. And on the fifth anniversary of El Salvador's Bitcoin legal tender law, the verdict is in from the IMF and from the data itself. Let's get into it.
AMD made a quiet acquisition that could reshape AI inference economics. They bought Taalas, a Toronto company with a genuinely radical idea. Instead of running a model on a general-purpose GPU that fetches weights from memory billions of times a second, Taalas etches the model weights directly into the silicon during manufacturing. The chip is the model. Their test chip, the HC1 built on TSMC's 6-nanometer process, ran Meta's Llama 3.1 8B at around 16,960 tokens per second. That's roughly 48 times faster than comparable Nvidia GPUs, according to their February reveal. The trick is eliminating the memory bandwidth bottleneck entirely. If the weights never leave the die, you skip the most expensive part of inference.
Now, the obvious catch. These chips are model-specific. A chip built for Llama 3.1 8B doesn't run Claude, doesn't run GPT, doesn't run the next Llama. Every new model needs new silicon. Taalas claims they can turn a trained model into a chip in about two months, which is fast for semiconductors but slow compared to shipping a software update.
AMD's play here is smart. They're not replacing their Instinct GPU line. They're adding a second tier. GPUs for training and experimentation, model-specific chips for high-volume production inference where you're running the same model billions of times. If you're a hyperscaler serving a specific frontier model at massive scale, per-token costs could drop by an order of magnitude. That's the pitch.
Meanwhile, UBS thinks AMD will eventually strike a foundry deal with Intel because TSMC capacity is permanently constrained. And Goldman Sachs data suggests Nvidia will still ship roughly 10 times more data center racks than AMD through 2028. So AMD is fighting on two fronts. Compete with Nvidia on volume, and try to leapfrog them on inference economics with specialized silicon. The Taalas bet is essentially that the future of inference isn't one chip serving many models. It's many chips, each fused to a single model. That's a very different world.
Riot Platforms just signed a 20-year, 9.1 billion dollar deal with Anthropic to lease 191 megawatts of compute capacity at their Rockdale, Texas campus. If Anthropic extends by two more five-year terms, the total revenue climbs to 16.1 billion. Combined with Riot's existing AMD partnership, they now have a two-tenant data center portfolio generating roughly 9.8 billion in contracted revenue. Riot is no longer really a Bitcoin miner. They're a power and real estate business.
This is the pattern now. Soluna Holdings just completed the buyout of Dorothy 1B in Silverton, Texas, taking full control of the 25-megawatt facility. Combined with their Briscoe wind farm acquisition and Dorothy 1A, they've built a fully integrated generation-to-compute campus, and they're explicitly bifurcating operations. Bitcoin mining moves to hosting only, while AI and high-performance computing take priority. Their development pipeline now exceeds 4.3 gigawatts.
The logic is brutal but simple. Post-halving Bitcoin mining margins are thin. Grid-connected power is scarce, and AI training clusters will pay premium prices for large blocks of near-term electricity. If you already own the substation, the land, the cooling infrastructure, and the interconnect agreement, you have something Anthropic and OpenAI cannot buy overnight at any price. ERCOT scrutiny of new power projects only makes existing capacity more valuable.
Contrast that with what's happening in Bolivia, where Alps Blockchain just fired up a decommissioned 127-megawatt natural gas plant in Cochabamba for pure Bitcoin mining, currently using 27 megawatts and generating 1.23 exahashes per second. Or the proposal from the Bitcoin Policy Institute to place miners near Ukraine's nuclear plants, potentially generating up to 1 billion dollars annually for reconstruction by absorbing surplus baseload power.
So you have a fork in the road. In markets with cheap stranded energy and weaker grid pricing, Bitcoin mining still works. In premium markets like Texas, the math tilts hard toward AI tenants. The Bitcoin maximalist reading is that this is fine. Mining migrates to the cheapest, most abandoned energy on earth, exactly where the incentive structure is supposed to send it. The market is doing its job.
Five years ago, El Salvador became the first country to make Bitcoin legal tender. Today the honest assessment is in, and it's not what the maximalist camp wanted to hear.
The three original goals were financial inclusion, cheaper remittances, and foreign investment. Let's take them one at a time. Financial inclusion. About 36 percent of Salvadoran adults had a bank account before the law. The Chivo wallet, backed by a 30 dollar Bitcoin signup bonus, did not meaningfully change that. Most users claimed the bonus and never transacted again. Remittances. This was the strongest case, given that remittances are about 24 percent of GDP and 98 percent come from the United States. The government projected 400 million dollars in annual savings. Actual result in the first half of 2026: crypto-channeled remittances totaled 35.4 million dollars, or 0.7 percent of total flows. Traditional intermediaries still handle over 84 percent. Foreign investment. No credible evidence the Bitcoin Law attracted meaningful FDI. A few small crypto initiatives, and that's it.
Then came the IMF. In 2025, El Salvador signed a 1.4 billion dollar Extended Fund Facility, and the conditions gutted the Bitcoin policy. Merchant acceptance became voluntary. Taxes must be paid in dollars. State accumulation stopped. And now the IMF is reportedly asking for names, wallet addresses, and BTC balances of all state-controlled legal entities using Chivo. That's a significant sovereignty concession dressed up as compliance.
But the government still holds around 7,400 to 7,700 BTC, worth roughly 480 to 500 million dollars. And with that concentration, a 10,000 dollar move in Bitcoin price swings the national budget by about 75 million. That's real fiscal exposure for a small economy.
Here's the honest read. As monetary policy, the experiment did not deliver its promised benefits. As branding, it worked. El Salvador became a global Bitcoin destination, Bukele became a Bitcoin celebrity, and the country proved a nation state can legally adopt Bitcoin without collapsing. The deeper lesson is about the ceiling. Any developing nation trying this hits the same wall the moment it needs multilateral credit. The IMF has veto power over sovereign Bitcoin strategy, and that's not changing soon. The next country that tries this needs a plan for what happens when the balance of payments turns against them. El Salvador didn't have one.
Cybersecurity is quietly becoming the most consequential real-world deployment of AI agents. OpenAI and AWS just made Daybreak Red and Daybreak Blue generally available to eligible customers on Amazon Bedrock. Daybreak Red is GPT-5.6 Cyber, focused on offensive vulnerability research. Daybreak Blue is GPT-5.6 Sol, focused on defense, detection engineering, and incident response. They run with zero-operator access at the chip level, data encrypted in transit and at rest, and customer data is not used for training. That last part matters. Enterprises will not deploy AI in their SOC if there's any chance sensitive incident data leaks into a foundation model.
On the vendor side, TENEX.ai's CEO Eric Foster is calling out the incumbents directly, describing legacy managed detection and response as, quote, a people business dressed up as technology. TENEX claims to analyze 100 percent of alerts with Tier 1 triage in under a minute, and deploys in seven days on top of Google SecOps or Microsoft Sentinel. BDO is layering agentic capabilities on top of Microsoft Sentinel with entity enrichment for Kerberoasting, AS-REProast, OAuth risk, and automated blast-radius analysis. In Korea, Estsecurity launched Alyac Agentic EDR integrated with LG's K-EXAONE model. Their numbers: 80 percent of analysis tasks automated, report writing 60 times faster, and a mandatory human in the loop for destructive actions like terminating processes.
But here's the piece that matters most long term. Docker, Snyk, and Keycard released Agent Baseline v1.0, a security framework specifically for enterprise agent deployment. Six outcomes: Discover, Constrain, Authorize, Observe, Validate, Respond. 35 controls total. The core idea is that AI agents are software with authority, and you cannot ship them into production without knowing exactly what they can touch, what credentials they hold, and how to yank them back when something goes wrong. Short-lived task-specific credentials. MicroVM sandboxes. A registry of every agent with owner and purpose. Stable run IDs to correlate intent with outcome.
This is the boring, plumbing-level work that will decide whether agentic AI actually gets deployed at scale in regulated industries. Not the model benchmarks. The governance layer. And the fact that the first serious framework came from Docker rather than a hyperscaler is interesting. The container people understand isolation better than the model people do.
One prediction to close on. The two biggest stories of the next 24 months are converging in the same physical location. Model-specific silicon like what AMD just bought from Taalas will need to run somewhere, and that somewhere is a data center built on power infrastructure that used to mine Bitcoin. The winners of the next cycle are the people who own substations, not the people who own models.