Google's Gemini 3.5 Pro is stuck in delay purgatory, missing its June launch because it can't hit internal coding benchmarks. Bitcoin is defending 62,500 dollars while futures open interest keeps climbing faster than spot demand can absorb. A Swedish company is about to list Europe's first Bitcoin-backed preferred shares with a 10 percent yield. And GitLab, Cognition, and Cast AI are all racing to sell governed AI coding agents to enterprises, with Forrester putting a 400 percent ROI number on it. Let's get into it.
Google's Gemini 3.5 Pro was supposed to launch in June. It's mid-July, and it's still not out. Bloomberg reported this week that the delay is because the model isn't meeting Google's own internal goals, especially on coding. That's the benchmark that matters right now, because coding is where Anthropic and OpenAI have been pulling ahead in enterprise deals.
The deeper story here isn't really about one model. It's about whether Google can actually ship. Reporting from current and former employees paints a picture of a company that can build frontier models in the lab but struggles to push them out the door across Search, YouTube, Maps, Android, Workspace, and Cloud all at once. Add strict internal AI safety review, disagreements about how much AI-generated code to trust, and heavy GPU demand from every team, and you get molasses.
Meanwhile, Google claims about 75 percent of its production code is now AI-generated, consolidated under a project called Antigravity. So internally they're eating their own dog food. But some researchers are leaving for Anthropic, and customer reception to Gemini 3.5 Flash has been mixed, with complaints that it's priced too high for the reasoning it delivers.
Google isn't sitting still on the enterprise side though. This week they announced Grounding with Parallel Web Search inside the Gemini Enterprise Agent Platform. It's a native web-grounding provider that lets Gemini pull real-time results with exact citations, with a zero data retention option for sensitive workloads. Billed through Google Cloud Marketplace. That's a smart move for agentic use cases where hallucinated sources are a dealbreaker.
And Gemini Omni Flash landed in Google Vids, letting Workspace users generate video with personal avatars that look and sound like them, watermarked with SynthID. Google Vids already has 7 million monthly users.
So the pattern is clear. Google is shipping incremental Gemini improvements at the product edges while the flagship model slips. In a market where competitors push updates in weeks, months of delay on the tentpole release is a real problem.
While Google fumbles the flagship, the AI coding agent market is quietly becoming a real enterprise category. Three announcements this week make that clear.
GitLab shipped 19.2 with what they call governed agentic automation. The headline features: Dependency Scanning Auto-Remediation, which opens merge requests to fix vulnerable dependencies and iterates in the same MR if the fix breaks the build. A Security Review Flow that catches business logic errors, race conditions, and authorization gaps that pattern scanners miss. And the GitLab Duo CLI going general availability, so agents run from the terminal with full project context. Everything routes through existing approvals and audit trails.
Alongside that, Forrester published a Total Economic Impact study, commissioned by GitLab, claiming 400 percent ROI and 7.5 million dollars net present value over three years, with payback under six months. The composite organization saw 80 percent faster developer onboarding, 75 percent faster code migrations, and a 20 percent gain in individual developer productivity worth 7.4 million dollars. Take vendor-commissioned numbers with the usual salt, but the direction is real.
Cognition's Devin platform just hit FedRAMP High In-Process status for the full platform, including Devin Cloud. U.S. Army, Navy, NASA JPL, Anduril, and Goldman Sachs are already customers. Their pitch is autonomous fleets of Devins doing security remediation, legacy code modernization, and cloud migration, with model-agnostic routing they claim delivers frontier performance at 35 percent lower cost.
And Cast AI shipped Kimchi Coding to general availability. It's an orchestration layer that routes each task to the cheapest capable model, mixing frontier models with self-hosted open-weight ones. They claim 2.5 times cost reduction versus a commercial-models-only baseline, with hard spend caps and full data sovereignty options including air-gapped deployment.
Meanwhile Replit is claiming 2.9 times more code output from engineers using their agents. Critics rightly point out there's no external audit, no clean denominators, no cost per accepted outcome. But review latency is flat, reversion rates are stable, and human PR review time is down 30 percent.
The common thread across all four: nobody is selling autocomplete anymore. They're selling orchestration, governance, audit logs, budget controls, and model routing. The AI coding agent is becoming plumbing.
Bitcoin is trading between 62,500 and 64,300 heading into the weekend. It rejected 65,500 twice in eight sessions, and Iran strike headlines have pulled it lower alongside U.S. stocks. But under the surface, the derivatives market is where the real story is.
Open interest in Bitcoin futures is around 48.9 billion dollars, up about 3.5 percent in two days. Spot ETF inflows in the same window were modest, roughly 289 million dollars. So leverage is expanding much faster than spot demand can absorb. That's a futures-led rally, not organic buying, and it's the kind of setup that gets ugly fast when direction flips.
CME Group cleared a record 459.2 billion dollars in notional across crypto futures and options in Q2. That's banks, asset managers, and pro desks using regulated venues for exposure and hedging, especially around spot ETF inventory. Notional doesn't tell you net positioning, but the infrastructure is clearly mature.
JPMorgan flagged an interesting split this week: spot Bitcoin ETFs turned to outflows, while CME and perpetual futures kept seeing positive inflows. Institutional money isn't leaving, it's just moving venues. And options desks are getting aggressive. CoinDesk reported large call spreads targeting 72,000 dollars by month end, timed to the Fed meeting. That's a specific, dated bet, not a vibe trade.
On Strategy, JPMorgan says cash reserves are up to 3 billion dollars, giving them roughly 20 months of preferred dividend coverage. That eases the concern that Saylor's operation would have to sell Bitcoin to fund STRC dividends. They plan to issue more STRC once it recovers to par at 100 dollars and use proceeds to buy more Bitcoin.
Short term, watch 62,500. Break below and it's range consolidation with support at 62,500 on the 20-day and 63,800. Break cleanly above 65,000 and the 72,000 call spreads start looking prescient. The bigger warning sign is the weekly close below the 200-week moving average earlier this cycle, the first since 2023. Reclaiming that takes weeks, not a candle.
Monday, July 20, Bitcoin Treasury Capital lists on Sweden's Spotlight Stock Market under the ticker BTC PREF A. It's Europe's first Bitcoin-backed preferred shares, and it's a genuinely interesting structure worth understanding.
Here's the setup. The company issued 195,078 Class A preference shares at 120 Swedish krona each, targeting about 23.4 million krona, roughly 2.5 million dollars. The preferred shares pay a fixed 12 krona annual dividend, distributed monthly. At the 120 krona issue price, that's a 10 percent yield. The entire dividend obligation is backed by the company's Bitcoin reserves, which currently sit at about 172 BTC, roughly 11 million dollars.
Why does this matter? Because it's a different template than the Saylor playbook. Strategy issues common stock and convertible debt against Bitcoin, and holders ride the volatility. BTC AB is offering a fixed-income instrument, with priority dividends, backed by Bitcoin as the underlying reserve. It's designed for institutional investors who want Bitcoin exposure with yield, not price beta.
The caution flag: the offering was not a blowout. Reporting suggests they raised about 12.2 million krona gross, subscription just over 50 percent. Nearly half the shares didn't sell even with a 10 percent yield. That tells you something about current appetite, or maybe about the size of the Swedish investor base that understands the product. Any discount below 120 krona on Monday pushes the effective yield above 10 percent.
Zoom out and the digital asset treasury sector overall is in a rough patch. Strategy authorized up to 1.25 billion dollars in Bitcoin sales this year, has already sold about 218 million to fund dividends and rebuild dollar reserves, and added a share repurchase program. Nakamoto and BitMine Immersion have also trimmed holdings. The mNAV metric, market value relative to net asset value, has fallen below 1 for many DATs, meaning shares trade at a discount to the crypto they hold. That's a broken business model.
BTC AB's preferred share approach is a hedge against that dynamic. Instead of relying on premium-to-NAV to raise capital, they're offering a yield product. If it works, expect copycats. If the 10 percent coverage ratio tightens under stress, expect scrutiny.
One prediction for the week ahead. If Bitcoin doesn't reclaim 65,000 before the Fed meeting, those 72,000 call spreads become the most expensive lottery ticket of the summer, and the leverage in futures is going to unwind harder than anyone is pricing in right now.