Bitcoin tagged $65,300 as a weak US jobs report cooled expectations for a September Fed rate hike. But the bigger story this weekend is BIP-110. Bitcoin has roughly 185 blocks left before enforcing nodes start rejecting non-signaling blocks, and developers are warning holders that if a minority chain appears, selling coins from the fork could get replayed on the main chain. Doing nothing is the safest move. Meanwhile, another Lightning infrastructure exploit hit BTCPay Server users running LND, with attackers stealing credentials that control Lightning wallets. Update immediately or take servers offline. In Washington, Senate Majority Leader Thune filed cloture on the CLARITY Act, teeing up a 60-vote showdown in September. And Trump Media walked away from its $6.4 billion CRO treasury deal with Crypto.com, refocusing on media and a pending fusion energy merger.
Let's talk about what's actually about to happen on the Bitcoin network this weekend. At block height 961,632, nodes enforcing BIP-110 will start rejecting blocks that don't signal for the proposal. That's roughly 185 blocks away as of this recording. Now, this doesn't mean Bitcoin automatically adopts anything. What it means is if miners on the non-signaling side keep producing blocks, and enforcing nodes reject them, you get a chain split. A minority chain appears alongside the main chain.
Here's the part every holder needs to understand. If you sell coins from the minority fork on any exchange, the signature you produce is valid on both chains. An attacker can grab that signed transaction and replay it on Bitcoin itself, draining your actual BTC. Until wallets and exchanges implement proper replay protection and can cleanly separate the chains, the developer consensus is simple: do nothing. Don't move coins. Don't try to claim free fork money. Don't sell either side.
This is a different situation than 2017's Bitcoin Cash fork, where replay protection was baked in from day one. Here it's not guaranteed. And it's happening against a backdrop where the network just went through an AI-driven security sprint that flagged 6,700 potential issues in 55 hours, though nobody has verified how many are actually real bugs versus noise. The combination of a live consensus event and an unverified flood of AI-generated findings is genuinely unusual territory for Bitcoin. Watch the block count. Don't touch anything until the dust settles.
Cybersecurity vendors have collectively decided that 2026 is the year the SOC goes agentic. In the past week alone: Pondurance launched Kanati, Arctic Wolf unveiled Aurora, SentinelOne expanded Purple AI, Securonix added governed AI agent detection, and Sysdig rolled out Secure AI. Five major announcements, same core pitch.
The numbers Pondurance is putting up are the ones worth pausing on. Kanati ingests over 60 terabytes of telemetry per day. It claims 90% AI-powered threat analysis with autonomous containment, under 2-minute average investigation time across all alerts, and an 80% reduction in false positives. Sub-second containment for high-confidence threats. And notably, it runs on Amazon Bedrock inside an isolated US-based AWS environment, with customer data walled off from model training.
SentinelOne's Purple AI is arguably the most battle-tested. It's been in production since June, running over 8,500 autonomous investigations per day, analyzing three times as many alerts as human analysts. That's not a demo. That's real production traffic.
The theme running through all of these is governance. Every vendor is emphasizing that humans define the boundaries. AI acts autonomously inside those boundaries, escalates outside them, and every action is traceable and reversible. This is the pattern that finally makes agentic AI defensible for regulated industries. Not full autonomy, but bounded autonomy with audit trails. If you run a SOC and you're not evaluating one of these platforms in the next six months, you're going to be outbid on both cost and response time by competitors who are.
Amazon quietly shipped something that changes the calculus for anyone building AI agents on AWS. Web Search on Amazon Bedrock is now generally available. It gives foundation models native, server-side access to a web index Amazon runs itself, containing billions of documents, plus a built-in knowledge graph for anchoring entities like authors and dates.
Why does this matter? Because until now, if you wanted your Bedrock model to have current web knowledge, you were wiring up third-party search APIs, managing API keys, building orchestration layers, dealing with data egress questions. Bedrock Web Search collapses all of that into a single parameter on an OpenAI-compatible API call. Configure credentials, point the OpenAI client at Bedrock, add the Web Search tool. Done. Zero data egress by default. Everything stays in AWS.
The retrieval approach is also smarter than the typical RAG pattern. Instead of dumping entire pages into context, it does semantic snippet extraction and returns only relevant passages with source URLs and titles. That means lower token costs and lower latency, with citations included in the response.
AWS also cut Bedrock inference prices for the GPT-5.6 family. Luna dropped 80%, now $0.20 per million input tokens and $1.20 per million output. Terra dropped 20%. And SageMaker's serverless customization now supports full fine-tuning on 25-plus open-source models, including gpt-oss, Gemma, Llama, and Qwen. Full parameter tuning, not just LoRA, with no infrastructure to manage.
The strategic read: AWS is not trying to win the frontier model race. It's trying to be the boring, compliant, cost-optimized place where enterprises actually deploy AI. Web search without vendor headaches, cheap inference, and fine-tuning without provisioning GPUs. That's a coherent story for CIOs.
Two stories this week frame where Bitcoin mining is actually heading, and neither of them is about hash rate. First, Texas. Governor Abbott ordered an audit of all data centers connecting to the ERCOT grid. Bernstein analysts say approved miners like IREN and Riot are largely insulated because their capacity is already contracted. But Cipher, Core Scientific, and CleanSpark could face pushback as ERCOT converts pipeline assets to grid-connected capacity. Cipher's stock dropped over 7% on the news.
The bigger picture is that Bitcoin miners have until 2027 to prove they deserve a seat on America's overloaded grid. US electricity demand is projected to rise through 2027, driven mostly by AI data centers, and grid operators are running out of patience with loads that just consume power without providing flexibility. ERCOT has flagged over 5,000 megawatts of large loads at risk of disconnection during faults. There have been at least 26 data center or crypto disconnections since 2023.
The bull case is miners becoming documented, dispatchable flexible loads that absorb renewable surplus and get paid for grid balancing services. The bear case is they fail to demonstrate reliable curtailment, and interconnection gets harder, contracts get pricier, and sites lose value unless they pivot to AI hosting.
Bitdeer is already making that pivot. They just signed a 16-year, $4.7 billion lease in Tydal, Norway for 121 megawatts of AI capacity, backed by $1.3 billion in letters of credit from JP Morgan. Bitdeer has also liquidated its Bitcoin treasury to zero to fund this expansion, which is the opposite direction from MARA, Riot, and CleanSpark, who are holding sizable BTC reserves. TeraWulf got FERC approval to take over Morgantown Power in Maryland, planning an initial 500 megawatt data center buildout with potential to reach a full gigawatt.
The miners that survive the next 18 months will look less like miners and more like flexible infrastructure operators. The ones that don't adapt lose their power contracts.
If you hold Bitcoin, keep your coins still through the BIP-110 activation window. If you run Lightning infrastructure, patch tonight. Everything else can wait until Monday.