TheChek News Flow

Wall Street’s boutiques bet on star bankers. Now they’re stuck with the bill

Investment banks like Evercore, Lazard and Moelis scooped up dealmakers ahead of an upswing that has yet to fully materialise

New York’s boutique investment banks insisted new rainmakers’ multimillion-dollar guaranteed payouts were a temporary phenomenon. The windfalls look like they are here to stay.

Banks, including Evercore, Moelis & Co and PJT Partners, have paid out at least 60 per cent of their net revenues to staff every year since 2022, the level long regarded as the industry benchmark.

After initially predicting the hangover from a post-pandemic hiring spree would last just two or three years — Lazard’s chief executive Peter Orszag said in 2024 he thought his bank’s compensation ratio would fall below 60 per cent the following year — boutiques are now warning investors they are in for a wait.

Evercore is “still a way from sub-60 per cent”, chief financial officer Tim LaLonde told investors in March. Lazard’s ratio hit 69.9 per cent in the first quarter of this year.

The boutiques’ costs shot up when they seized on a post-pandemic downturn in dealmaking as an opportunity to snap up dozens of senior bankers shed by Wall Street’s bulge-bracket players before the business cycle turned again.

Recruits take time to build a book of business — not least because of lengthy non-compete clauses.

“[We] made the observation at the time [of our IPO] that it was quite difficult for any new partner . . . to generate any revenues of consequence in the first two years,” Paul Taubman, founder and chief executive of PJT Partners, recently told analysts.

To persuade them to take the risk on a smaller firm, the boutiques offered the bankers guaranteed minimum payouts for the first two years followed by a promise of a 25-30 per cent cut of their deal revenues.

In theory, the investment should pay off when dealmaking rebounds and a bank is fully staffed to capture the upswing. And in some cases, fees have soared. From 2022 to 2025, PJT’s investment banking advisory revenue jumped more than 80 per cent; at Evercore they rose 36 per cent. Total revenues at Moelis increased by almost 60 per cent in that time.

But rather than the post-pandemic hiring spree being a one-off event, the boutiques have continued to add expensive new recruits as mid-market dealmaking has struggled and they have pushed into faster-growing areas such as private credit, AI, power and energy, tech and sports.

Lazard added 28 senior investment bankers in 2025 against a target of 10 to 15, pushing up its compensation ratio. It has since abandoned any particular target date for hitting the 60 per cent threshold.

The fact that these banks have continued to hire expensive senior bankers despite stubbornly high compensation ratios would typically be a sign of confidence that higher revenues are just around the corner.

“Why is it so expensive to hire right now? Because people are optimistic about their businesses and so they want to pay up for talent,” said Devin Ryan, equity research analyst at Citizens JMP.

“We’re in a very hard market for recruiting right now because of the revenue potential. I think people feel like this is quite bullish for this business.”

But it could also indicate a more enduring shift in boutiques’ models, where a higher share of revenues is diverted to banker pay instead of shareholder payouts like dividends and share buybacks.

Boutiques’ shares have underperformed the S&P 500 this year — as well as their larger peers, which have benefited from soaring revenues in their trading units.

There are worries that while megadeals are still being struck, the midsized transactions that are banks’ bread-and-butter have slowed because of high financing costs, a stagnant private equity industry, and geopolitical uncertainty.

“When [the boutiques] were drawing up the plans in 2024 for hiring, they expected 2026 would look better than it does,” said Brennan Hawken, senior equity research analyst at BMO Capital Markets. At Lazard, investment banking advisory fees climbed just 10 per cent between 2022 and 2025.

Executives are looking for other ways to cut costs, including by using AI to save on junior bankers.

Lazard’s Orszag has been most enthusiastic about using automation to enhance client work and cut the number of junior staff required to support a senior banker, something he has described as the “total associate equivalent per managing director ratio”.

“It’s a little bit of a back to the future on the team size, because it used to be that teams were smaller,” Orszag told an industry conference last month. “As specialisation took over, team sizes got bigger. We’re going to be going back to smaller team sizes enabled with these new [AI] tools.”

But others are more sceptical. The chief executives of Moelis and Evercore have said publicly they did not immediately see the opportunity to reduce staff. JPMorgan Chase chief Jamie Dimon said in May that he thought banks could use AI to “create temporary margin, but not permanent margin”, as competition eroded the gains in profitability.

There is a chance that if midmarket dealmaking remains thin for much longer, banks may have to start culling rainmakers. Perella Weinberg recently cut a tenth of its workforce.

One longtime bank executive in charge of pay decisions said banks found reasons to justify high packages for star bankers whatever the stage of the business cycle.

“It’s a pretty slim-margin business,” the executive said. “They do a lot of tap-dancing around ‘investing in growth’ or ‘retaining employees’, but the bottom line is the ratio ends up high in good times or bad.”

Exit-entry system is forcing some passengers to wait on the tarmac in intense heat, Alexander Zinell says

The head of 14 Greek airports including Corfu has warned that Europe’s new smart border system will not work in its current form and needs to be completely “overhauled”. 

Alexander Zinell, chief executive of Fraport Greece, said the EU’s new entry-exit system had “fundamental flaws” by forcing passengers to carry out lengthy checks at airports, leaving some waiting on the tarmac in intense heat.

Greek police have suspended all biometric checks on UK citizens that make up the vast majority of the non-EU tourists that pass through his airports, which include Rhodes, Crete and Mykonos. The exemption, which will last for the summer, is the only thing preventing it from “collapse”, he said. 

“These are just temporary fixes, the system needs to be overhauled,” he told the FT. “It needs a new version, an update, and probably a reconfiguration in order to allow people to register before they fly, before they get on a plane, before they go to the airport.”

The EES began rolling out last October but has led to long queues and disruption since it was fully switched on in April.

Passengers have queued for hours, there have been widespread technical problems with automated booths, while Zinell said “many” planes had been forced to leave behind passengers who were stuck waiting to be processed. 

Queueing at some of its sites has been so bad, gazebos have been erected to provide shelter from the sun to people waiting on the tarmac to be processed, he said. The airports have designated “queue walkers” who fast-track vulnerable people from the lines. “It is very unpleasant for passengers, and even dangerous,” Zinell said.

European Commission president Ursula von der Leyen conceded last week there were “technical problems” with the system, which includes automated booths where passengers provide fingerprints and have a photo taken. 

Zinell said his airports — which process hardly any passengers in winter and millions during July — are not equipped for long waiting times. 

“Our facilities are designed for typical standard processing times. Once you extend these times, you get queues. It’s very simple,” he said. “If you quadruple the average time, it means that you need four times the space.” 

He said it was “absurd” for airports to build more buildings for an electronic system that can in theory be done online. 

“The UK, Canada, US, they all have similar systems. I think most countries nowadays require you to preregister before you fly. And that makes sense.” 

Plans by the EU to roll out a US-style pre-authorised travel system have been delayed until next year given the current difficulties with the EES and pending more technical tests.

“We fully support [the EES] at the airport because it’s a matter of security,” Zinell said. “But the system needs to be designed in a way that fits into processes at the airport. Unless this is changed, unless the process time, through whatever means, is brought down significantly, we will face these issues.”

Under existing rules agreed by member states, the exemptions — such as those for UK citizens — will not be allowed from September. 

Zinell said this approach is “denying reality” by refusing to show any leeway. “I am more worried now about what’s going to happen after the summer and next summer,” he said. 

“If we had a situation where nothing changes and people had to wait for four hours, it’s just not workable.”

The European Commission has so far rebuffed calls to fully suspend the system or to extend the exemptions.

“The system doesn’t work if you have full suspension in one or the other entry or exit points . . . you constantly need to reconcile entry and exit,” a commission official said. Otherwise, the official explained, people whose exit is not registered could be considered overstayers and barred from entering the next time.

An extension of the current exemptions, which allow border guards to only register personal details without image or fingerprints, was not planned at the moment, the official added.

“For the time being, we have no requests [from EU nations], also no objective grounds to start looking at a possible extension post September.”

The days of Nvidia’s unparalleled market dominance aren’t over, but challengers and choices are arising from all directions.

ZML, a hot French AI startup endorsed by Turing Award winner Yann LeCun, has released inference-performance software that allows a variety of open source large language models to run on a variety of chips — including Nvidia’s, AMD’s, Google’s TPU, Apple Metal, and Intel Arc.

With ZML/LLMD, the newly launched LLM inference server, the company’s ambition is to break existing silos and make different chips available for AI use cases at their maximum available speed, and sometimes faster, ZML founder Steeve Morin told TechCrunch.

As AI becomes integrated into our work and everyday lives, optimizing inference — aka, the processing of prompts — has been outpacing model training in importance, but often feels patchy behind the scenes, with software and architecture barriers that lead to vendor lock-in, Morin said.

The promise of achieving peak performance across a variety of chips is a technological feat, but it could also be a market disruptor, amid mounting fears over AI-related costs.

ZML hopes to provide enterprises and clouds with the option to use a mix of chips, some of which might be less costly or consume less energy. “The idea is to give people back the power to create their own system and achieve real efficiency gains that allow [AI] to be disseminated,” Morin said.

Such a software assist may help novel AI chipmakers, many of which happen to be from Europe, Morin observed, citing Axelera, Fractile, Kalray, OLIX, Q.ANT, SiPearl, SpiNNcloud, and VSORA. But more than their region of origin, what matters to him is that ZML can work with them on “things that haven’t been done before anywhere in the world.”

That doesn’t mean Morin is bearish on Nvidia. He’s not, in part because of its existing supply. He told TechCrunch that ZML has a good relationship with the AI chip giant, which has been gearing up for the rise of inference.

Inference has been an area of such intense investment that the trend has been hailed the “inference gold rush.” So ZML has competition such as Baseten, recently valued at $13 billion; Inferact, from the creators of open source project vLLM; as well as RadixArk, the commercial company behind SGLang.

Both vLLM and SGLang partially compete with LLMD, but Morin’s ambitions for ZML cover a broader spectrum. “We have reached the point where we are co-designing silicon,” he said. He further credited ZML’s lean team of 20 people as the reason why the Paris-based startup has been able to move fast, with more releases in the plans.

It also helped that this small team is well funded for its size. Thanks to his track record as VP of engineering of Zenly, which Snapchat acquired for nine figures in 2017, Morin raised $20 million from venture firms including Harry Stebbings’ 20VC, >commit, AALVC, Drysdale Ventures, Xavier Niel’s Kima Ventures, Kindred Capital, LocalGlobe, and Puzzle Ventures.

Unlike ZML’s first public project, the inference-focused ML framework released in 2024 and updated in March, ZML/LLMD is not open source. But it is launching as a free product with the goal of learning about usage. “I’d rather measure and [then generate revenue] where it is most effective without hindering my growth stupidly because I have been too greedy from the get-go,” Morin said.

It is too early to tell when ZML/LLMD might become a paid product, and what its adoption will look like. But the startup’s cap table confirms that other founders are paying attention, including Dagger and Docker founder Solomon Hykes, Clément Delangue and Julien Chaumond from Hugging Face, as well LeCun, now with AMI Labs. This also builds the case that Europe’s AI startups can now build from home. “I couldn’t do ZML anywhere but in Paris,” Morin said.

D’après la start-up française, ce modèle est conçu pour «la navigation robotique, permettant aux robots de se déplacer de manière autonome dans des environnements complexes, notamment des bureaux, des bâtiments résidentiels et commerciaux, ainsi que des espaces extérieurs».

L'entreprise française d'intelligence artificielle (IA) Mistral, champion européen du secteur, a dévoilé mercredi son premier modèle d'IA spécialement conçu pour être utilisé sur un robot et entend se développer dans ce domaine. L'annonce de ce nouveau modèle, baptisé Robostral Navigate, est accompagnée d'une vidéo montrant un robot humanoïde sur roues, qui se déplace dans un bureau en suivant des instructions.

D'après la start-up française, ce modèle est conçu pour «la navigation robotique, permettant aux robots de se déplacer de manière autonome dans des environnements complexes, notamment des bureaux, des bâtiments résidentiels et commerciaux, ainsi que des espaces extérieurs». Il comprend des informations captées via les images d'une caméra intégrée au robot et des instructions en «langage naturel», a précisé Mistral, qui donne un exemple de consigne: «Quitte le hall, traverse le couloir, entre dans la salle de stockage et arrête-toi face à la deuxième étagère».

Mistral développe ses activités

Il peut fonctionner avec des robots munis de roues, de pattes ou volants, quelle que soit leur taille, a ajouté la start-up. Signe que l'entreprise, lancée en 2023, entend poursuivre ses travaux sur ce segment, elle écrit sur son site qu'elle veut recruter des ingénieurs et chercheurs pour étoffer son équipe spécialisée dans la robotique.

En parallèle, Mistral développe ses activités dans le domaine de la défense et de l'industrie. En mai, la start-up a officialisé la signature d'un contrat avec Airbus. Côté cybersécurité, le dirigeant de Mistral, Arthur Mensch, avait confirmé que sa société travaillait sur un produit spécifique pour les entreprises, après que le géant américain Anthropic a dévoilé l'existence de Mythos, modèle réputé hyperpuissant dans la détection de vulnérabilités.

« Le marché de la voix grandit à très grande vitesse » : Nvidia entre au capital de la start-up d’IA française Gradium

Le géant américain investit près de 30 millions de dollars dans ce spécialiste de l’IA vocale, convaincu par le potentiel de ce marché.

Nvidia s’invite au capital de la start-up Gradium, spécialisée dans l’intelligence artificielle vocale. Le géant américain, numéro 1 mondial des processeurs graphiques, investit près de 30 millions de dollars dans la jeune pousse. Il a déjà accompagné ces derniers mois plusieurs pépites hexagonales (AMI Labs, Alice & Bob, Scintil Photonics).

Moins d’un an après son lancement, Gradium franchit ainsi le cap des 100 millions de dollars de financement. « C’est un signe de confiance dans la trajectoire d’expansion de Gradium et dans le potentiel du marché de la voix, qui grandit à très grande vitesse », se félicite Neil Zeghidour, l’un de ses trois cofondateurs, qui comptait déjà comme actionnaires plusieurs fonds (FirstMark Capital, Eurazeo, Korelya Capital, Amplify Partners), ainsi que des bienfaiteurs du laboratoire Kyutai (Eric Schmidt, Xavier Niel, Rodolphe Saadé). La start-up profitera des fonds levés pour s’installer au cœur de la Silicon Valley. « Nous voulons être un acteur de classe mondiale et le leader dans ce domaine », poursuit cet ancien de Meta et de Google, prêt à se mesurer aux géants de la tech.

Issue du laboratoire de recherche Kyutai

Gradium est la première entreprise commerciale issue du laboratoire de recherche Kyutai, fondé en 2023 par Xavier Niel (Iliad), Rodolphe Saadé (CMA CGM), et Eric Schmidt (ex-directeur général de Google). Un laboratoire qui s’est d’emblée fait remarquer pour son innovation. Six mois après sa création, Kyutai rendait public son premier projet d’ampleur, Moshi, une intelligence artificielle capable de converser en temps réel, voire d’interrompre son interlocuteur.

Forte de ce savoir-faire, Gradium a déjà engrangé plusieurs milliers de clients, PME, ETI ou grandes entreprises, dans les domaines de la santé, de la banque, du commerce, du jeu vidéo ou de l’apprentissage des langues. « Nos clients utilisent nos modèles essentiellement pour les interactions au téléphone (support client, prise de rendez-vous, suivi de commandes, réservation d’un service…), détaille son fondateur. Les IA vocales présentent l’avantage de ne jamais s’agacer, d’être joignable à toute heure du jour ou de la nuit et de ne pas faire attendre. » Il est convaincu que la voix deviendra l’interface principale entre les individus et les systèmes intelligents.

Les interactions au téléphone, une priorité

Gradium a concentré ses efforts sur la création de modèles adaptés à ces interactions téléphoniques, le domaine en plus forte expansion. « Nous cherchons à créer des interactions naturelles, précise Neil Zeghidour, dont l’entreprise travaille avec des doubleurs qui prêtent leur voix à ses modèles. Il est important d’entendre une voix qui respire, des hésitations… tout en arrivant à une prononciation de pointe pour les contenus complexes comme les acronymes, les adresses e-mail ou les numéros de téléphone. »

À son actif, la start-up, qui débarque dans un marché très concurrentiel préempté par les géants de l’IA, a cinq modèles de synthèse et de reconnaissance vocale qui sont mis à jour toutes les deux à trois semaines. « Un des facteurs de notre succès est notre capacité à adapter très vite nos modèles aux attentes de nos clients, en fonction de leur retour. Dans un secteur qui progresse très vite, cette réactivité est une façon de nous imposer face aux acteurs plus gros », précise Neil Zeghidour. La jeune pousse, qui a recruté plusieurs anciens des Big Tech, se targue d’avoir « créé une équipe d’élite », capable de lancer un modèle en trois à six semaines, contre trois à six mois en moyenne pour les acteurs du marché.

Modèle de traduction en temps réel

Gradium a déployé récemment son premier modèle de traduction en temps réel, qui peut être intégré à des visioconférences ou à des livestreams. « Nous arrivons à un équilibre entre vitesse et précision de la traduction meilleur que les modèles d’OpenAI ou de Google », ajoute le jeune dirigeant. Dans les prochaines semaines, des applications dans les médias devraient se concrétiser, à savoir des « expériences de contenus personnalisés autour du podcast ».

Gradium met en avant l’accessibilité de ses modèles, jusqu’à trois fois moins chers que ceux de concurrents comme ElevenLabs. S’il facture essentiellement ses IA en fonction du volume d’usage, il a également introduit le principe d’une licence au tarif fixe. Un modèle plus rentable notamment pour les développeurs de jeux vidéo gratuits. « Plutôt que de tourner dans le cloud, ces modèles tournent directement sur le smartphone, précise Neil Zeghidour. Nous y croyons beaucoup pour la création de médias et d’applis centrés autour de la voix. »

Ambitions à l’international

Nvidia s’est avéré être un partenaire naturel pour Gradium. Des liens scientifiques unissaient déjà le géant américain et la start-up française. Nvidia a bâti une IA vocale à partir des modèles open source de Gradium. De son côté, la jeune pousse utilise les processeurs graphiques de Nvidia et figure parmi ses premiers testeurs. « L’utilisation de nos modèles sur leurs machines permet d’avoir des interactions rapides et le sentiment de parler à un humain », explique Neil Zeghidour.

Le financement du géant américain, qui va lui permettre d’accélérer la recherche en IA et le développement produit, devrait aussi lui donner des ailes à l’international, à commencer par les États-Unis. Gradium a commencé à recruter les membres de son futur bureau californien. L’un des cofondateurs pourrait à terme s’y installer pour accélérer le développement. C’est dans la Silicon Valley que sont situés la plupart de ses clients potentiels, souvent issus de la tech. Gradium lorgne déjà sur l’Asie. « Un marché que nous souhaitons ouvrir assez vite, en proposant des modèles conversationnels dans les langues locales », précise Neil Zeghidour.