HOST AFriday means we look at what we said would happen. And uh, this one is a little weird.
HOST BYeah. Weird in the way a clean whiteboard is weird after a messy week. We have no resolved wins and no resolved losses in the ledger.
HOST ASo this is less a victory lap and more a calibration meeting with microphones.
HOST BWhich is honestly useful. If nothing has resolved, we can still ask whether our open predictions are aging well or just sitting there looking confident.
HOST AAnd we should say it plainly: this week we don't have many resolved ones to roast.
HOST BExactly. So let’s do the honest thing and inspect the live bets.
HOST AFirst one: Google will turn Gemma 4 into a Chrome-side agent. Confidence 55%.
HOST BThat’s the one where we’re basically saying Google won’t just ship a model, it’ll ship an execution layer for browser work.
HOST AThe signal we liked was the model momentum plus the broader agentic shift. Google DeepMind has Gemma 4 in the news cycle, and the browser is the obvious surface if you want consumer task completion.
HOST BAnd the graph context wasn’t subtle: Google had high recent momentum. That said, this is still a quarter-ahead prediction, not a next-week call.
HOST ARight. If we’re honest, this one feels directionally right, but the timing could still embarrass us.
HOST BThat’s the risk. Google can be maddeningly deliberate. We may be reading the market’s appetite for browser agents correctly while being early on the packaging.
HOST AStill, of the open items, this one feels like it has the cleanest narrative path. Chrome is already the natural place for agentic browsing workflows.
HOST BAgreed. Not a win yet, just not a bad bet.
HOST ASecond: Microsoft will split Copilot agent billing from M365.
HOST BThis one feels more financially inevitable than theatrically exciting.
HOST AWhich is usually a good sign for a prediction engine, if we don’t get too cute about it.
HOST BThe signal was capacity pressure and monetization pressure. Microsoft has every incentive to stop pretending agentic workflows can live forever inside one tidy bundle.
HOST ABut here’s our self-own: we may have over-weighted the logic of the business model and under-weighted Microsoft’s love of delaying awkward pricing conversations.
HOST BYeah, that’s fair. We can be a little too tidy in our head. Real enterprise packaging often moves slower than the economics suggest.
HOST ASo if this doesn’t land soon, it won’t necessarily mean we were wrong on direction. It may just mean the company is stretching the bundling phase longer than we expected.
HOST BThat’s the calibration point. We’re not betting on whether Microsoft eventually monetizes agents separately. We’re betting on whether it does it within the quarter.
HOST AAnd that timing edge is exactly where we can get overconfident.
HOST BSpeaking of overconfidence, let’s talk Google again: Google will split TPU pricing for inference by Q3 2026.
HOST AThis one is a pure economics story. We saw TPU sales opening to select customers, plus hardware-commercialization pressure, and we thought inference-heavy workloads would force a separate price lane.
HOST BI still think the logic is strong. But the lab’s weak spot here is assuming that ‘strong logic’ becomes visible pricing quickly.
HOST AYep. We can map the incentive structure, but pricing teams move like they’re carrying furniture up stairs.
HOST BThe good news is the signal cluster is real: the hardware side, the cloud side, and the need to make inference economics legible to developers all point the same way.
HOST AThe bad news is that none of that guarantees a clean SKU announcement on our schedule.
HOST BSo if we’re wrong here, it’ll probably be a timing miss, not a thesis miss.
HOST ANow the more chaotic one: MCP security vendors will pivot from connectors to policy enforcement.
HOST BThis one feels like the market is already half there. You can only sell connector discovery for so long before customers ask, ‘Okay, but who’s approving the tools?’
HOST AThe reasoning cited MCP pressure rising alongside agentic workflows, and that still looks alive. Claude Code uses MCP directly, Claude Agent is surging, and multi-agent systems are still screaming upward.
HOST BWhat we’re watching for is whether vendors stop marketing the plumbing and start marketing the guardrails.
HOST AThat’s the part I like. Security markets usually follow the pain, not the elegance.
HOST BAnd if we’re honest, this is one of the least glamorous predictions we made, which is often a good sign.
HOST ABecause it’s less about buzz and more about what enterprises will actually pay for.
HOST BStill, we should not act like the evidence is already in. This is a structural shift we expect, not a confirmed pivot.
HOST ALast one: DeepSeek will trigger a pricing reset in Asian developer tooling.
HOST BThis one is interesting because the signal is more market-behavioral than product-specific.
HOST AWe keyed off DeepSeek’s $45B valuation in its first VC round and the fact that sentiment is rising even while it competes with the big frontier players.
HOST BThe thesis is that local wrappers and middleware stop selling ‘best model access’ and start selling efficiency, cost, and fit.
HOST AIt’s a neat thesis. Maybe a little too neat.
HOST BThat’s the honest worry. We may be projecting a broad pricing reset from a single very loud signal.
HOST ABut I don’t want to undersell the pattern either. When cost efficiency starts to dominate the conversation, downstream tooling usually feels it first.
HOST BTrue. This one just needs more proof than vibes. If two or more Asia-facing developer products repackage around DeepSeek-style economics, then we’ll know we read the room correctly.
HOST ASo, no resolved wins, no resolved losses. That’s not exciting radio, but it is the truth.
HOST BAnd the truth is useful: our current batch is mostly about timing and packaging, not about wild thesis errors.
HOST AWhat did we learn from that?
HOST BProbably that we’re decent at spotting the direction of travel, but we still get twitchy about the exact quarter.
HOST AYeah. We’re a little too eager to turn ‘the market is moving this way’ into ‘it will show up in the next 90 days.’
HOST BThat’s our recurring bias: we’re better at structure than cadence.
HOST AAnd maybe a second bias: we trust visible momentum signals a little more than we should when the commercial layer is slow.
HOST BEspecially with Google and Microsoft. Those companies can agree with the thesis and still make us wait for the productized version.
HOST AThe upside is that our open predictions are mostly coherent. The downside is that coherence can hide impatience.
HOST BWhich is a polite way of saying we sometimes get cocky about timing.
HOST ABad take? Not this week, because nothing resolved badly.
HOST BBut plenty to monitor badly if we get lazy.
HOST AFor next week, I’d add a cautious meta-prediction: at least one of the Google or Microsoft packaging bets will get more concrete, but probably not all the way to a clean commercial split.
HOST BThat feels fair. The lab should keep pretending to be humble until the evidence forces us otherwise.
HOST ATime to look in the mirror, then keep the whiteboard up.
HOST BExactly. Friday check-in complete. No trophies, no autopsies, just a few predictions still breathing.