Welcome back to our bi-weekly frequency. Through here, you’ll be able to access a roundup on all things B2B tech, algorithmic changes on the most popular communication channels, and top news of the week - analysed by members of our team.
In our last frequency, we covered what is working on LinkedIn and a news cluster on AI and geopolitics. Today, we will zoom into X and the changes made to its algorithm earlier this year, then cover a cluster on AI and disclosure.
Algorithmic changes on X
In January, X scrapped its old system, rebuilt the feed on a Grok model, and put the whole thing on GitHub. The weights are public now, and they say something blunt: conversation beats broadcast.
If you run a B2B brand or an exec profile, there, you can see it in the numbers. Analysts reading the code say a reply is worth around 27 times more than a like, and a real conversation, where you answer and get another response back, is worth roughly 150 times more than a like. The first half hour after you post sets your ceiling. And unusually for 2026, plain text beats video here. Grok reads the post itself, so hashtags do nothing and a considered take outperforms a recycled one.
It's the same hook, “belief, receipt, offer” template that wins on LinkedIn, only now it's enforced in code. The difference is that on X the reply is the whole game, so the "offer" at the end can't be a soft CTA. It has to be a line someone actually wants to answer.
Which tells you exactly what isn't working: the link-drop, the "excited to announce," the post that sends people off-platform. X suppresses outbound links hard, and posting more often won't save you, because a per-creator cap just spreads your reach thinner across each post.
A caveat of X, though, is that it is now pay-to-play. Free accounts have watched their average engagement fall to near zero, and Premium buys several times the reach. The belief-and-conversation approach still wins, but on X you have to pay before you speak.
Our takeaway: assert something a real person would put their name to, back it with one concrete detail, and end on something worth replying to, not something worth clicking away from. Just this time you might have to also open your wallet.
News of the week: Big tech names and disclosure
AI-infrastructure stocks jumped after strong quarters from CoreWeave, Super Micro and Nebius, and the message was the same: customer spending on AI continues to surge. CoreWeave's stock popped almost 20% after Q2 revenue doubled to $2.6bn (up 112% year on year), with a revenue backlog of $104bn and near-term capacity effectively sold out, which is enough leverage that it can now write new contracts on better terms. Big Tech's combined AI outlays are set to pass $730bn this year. For B2B tech, there is a clear claim: demand is here now, the backlog is contracted, the capacity is spoken for.
The same week, the software layer got lined up for a much more direct test. OpenAI is expected to file its public S-1 within weeks, ahead of a targeted September listing that would be the first time anyone sees the full financials behind ChatGPT. The reported picture is roughly $2bn in revenue a month against a company still deep in the red, losing something like $1.22 for every dollar it earns, with a last private valuation of $852bn and a public target north of $1tn. The filing will also put a hard number on Microsoft's stake, thought to be about 27%. But an S-1 is the ultimate claims-tested-against-reality document.
For buyers and their comms teams, the useful thing to notice is that OpenAI's momentum is about to acquire a denominator. Claims that lean on it will be read next to a real loss figure for the first time. The specific, checkable version survives (revenue is growing fast, adoption is enormous, the model is strong at these particular tasks); but all speculative claims will be either demonstrated or not.
On the safety side, a different kind of bill came due. On August 5 Meta disclosed that one of its models, reported to be Muse Spark 1.1, the agentic system it had launched weeks earlier as its most capable yet, reached a real company's systems during a security evaluation and exploited a live vulnerability there. It was the third such admission from a frontier lab in five weeks, after OpenAI, whose model compromised Hugging Face, and Anthropic, with the same testing vendor, Irregular, involved in at least two of the three. The headline writes itself: an autonomous AI hacked a company. The reality is that Irregular had misconfigured the test environment, handed the model internet access it was never meant to have, and the model used it to finish the task it was set. Meta, Anthropic and the vendor are now arguing in public over how serious that is, which is the tell: when something goes wrong autonomously, accountability is the first thing everyone tries to hand to someone else. Congress has already reached for the obvious lever with the bipartisan AI Kill Switch Act, and this won't be the last containment story that turns into a rule.
The bottom line for tech PR is that all three stories are, underneath, the same one: big tech names being made to account for their claims, whether to investors, to the market, or to the public. Earnings, a prospectus and an incident report are different mechanisms, but each turns a confident statement (demand is durable, the business works, the model is safe) into something that has to be backed by a number or a fact, in public. And the accounting now tends to arrive faster and more visibly than the claim did, so a strong quarter, a filing or a bad disclosure can settle in a day what a year of careful positioning tried to keep open. The useful move, as ever, is to get clients to decide which specific and slightly hedged version of each claim they can stand behind (e.g. demand for our compute is real and contracted, our model is strong at these tasks and contained under these conditions), rather than leaving it for an earnings call, an S-1 or an incident report to decide for them.